My Conception of a Swarm of Agents. Courtesy of Cocoa Yeo (2008).
what's new
ICML 2021 Expert Reviewer ICML 2020 Top 33% Reviewer ICML 2019 Top 5% Reviewer


Invited to serve as a World Economic Forum’s Global Future Councils Fellow for the Council on the Future of Artificial Intelligence and Robotics, Sep 2016  Jun 2018


Keynote speaker at 20th IEEE/WIC/ACM International Conference on Web Intelligence and Intelligent Agent Technology, 1417 Dec 2021


Organizing CoChair for NeurIPS 2021 Workshop on Closing the Gap between Academia and Industry in Federated Learning: Challenges on Privacy, Fairness, Robustness, Personalization and Data Ownership


Keynote speaker at 2nd
International Symposium on MultiRobot and MultiAgent Systems, 2223 Aug 2019


IEEE RAS Distinguished Lecturer for the
IEEE RAS Technical Committee on MultiRobot Systems, Mar 2019


Recipient of Faculty Teaching Excellence Award
Aug 2017  Jul 2018


Our 6 NeurIPS 2023, 3 ICML 2023, 3 ICLR 2023, 2 AISTATS 2023, AAAI 2023, 1 MLJ, 1 AIJ submissions are accepted!


AI Singapore Research Programme : Toward Trustable Modelcentric Sharing for Collaborative Machine Learning,
S$8,401,002.40, Apr 2021  Mar 2025


RIE2020 AME Programmatic Fund : Learning with Less Data,
SGD $1,218,600, Apr 2021  Mar 2024


DSTA Project Agreement : Tactics Discovery and Recommendation,
SGD $1,143,120, Jul 2021  Aug 2023


MOE AcRF Tier 1 Reimagine Research Scheme Funding : Scalable AI Phenome Platform towards FastForward Plant Breeding (Machine Learning)
SGD $348,600, Mar 2021  Mar 2024


RIE2020 AME IAFPP :
High Performance Precision Agriculture (HiPPA) System, SGD $1,197,960, Mar 2020  Feb 2024


Invited to serve as
area chair of ICML 2024, ICLR 2023, 2024, AISTATS 2023, 2024, ECAI 2023, AAAI 2022, 2024, RSS 2022, CoRL 2020, associate editors of IROS 2012, 20202023 & ICRA 2011, 20202023, IEEE RAL
senior program committee members of AAAI 2019, AAMAS 20182019, 2023, IJCAI 2015, 20202022, 2024, ECAI 2020, ICRA 2022
program committee members of AAAI 2010, 20162018, 20202021, 2023, IJCAI 2011, 2015, 2017, 2019, UAI 20212023 AAMAS 20112014, 2016, 20212022, RSS 2014, 2018, 2020,
CVPR 2021,
ICAPS 20102012, 20182019, 20212022, and
reviewer of NeurIPS 20132016, 20182023, ICML 20192022, AISTATS 20192022, ICLR 20192021


Students interested to join MapleCG, click here for more info


members
I am looking for talented undergraduate and graduate students in NUS to join my MapleCG research group.
If you are really excited and motivated to be involved in novel research in the fields of
artificial intelligence, planning under uncertainty (i.e., decisiontheoretic, informationtheoretic),
robotics, multiagent systems (i.e., multiagent coordination, planning, and learning), game theory,
statistical machine learning, optimization, and/or swarm intelligence, please email me and we can set
up a time for discussion. Please also take some time to view our research projects.
I am currently advising the following students and research staff:

Zhao, Zitong
Ph.D.
B.Sc. in Computer Science and B.Sc. in Data Science > University of Michigan  Ann Arbor
Research Interests: large language models


Niu, Xinyuan
Ph.D. (Coadvised with ChuanSheng Foo, A*STAR)
Recipient of A*STAR Graduate Scholarship (AGS)
B.Eng. in Mechanical Engineering (Hons. 1st Class) with Second Major in Innovation & Design and Minor in Computer Science > National University of Singapore
M.Eng. in Mechanical Engineering > National University of Singapore
Research Interests: robotics


Chen, Zhi Liang
Ph.D. (Coadvised with ChuanSheng Foo, A*STAR)
Recipient of A*STAR Computing and Information Science Scholarship (ACIS)
B.Sc. in Computer Science (Hons. 1st Class) and B.Sc. in Applied Mathematics (Hons. 1st Class) > National University of Singapore
Research Interests: automated machine learning


Wang, Jingtan
Ph.D. (Coadvised with ChuanSheng Foo, A*STAR)
Recipient of A*STAR Computing and Information Science Scholarship (ACIS)
B.Eng. in Computer Science with Minor in Mathematics > Nanyang Technological University
Research Interests: large language models


Chen, Jiangwei
Ph.D. (Coadvised with ChuanSheng Foo, A*STAR)
Recipient of A*STAR Computing and Information Science Scholarship (ACIS)
B.Sc. in Computer Science (Hons. 1st Class) with Major in Mathematics > National University of Singapore
Research Interests: collaborative machine learning, machine unlearning


Apivich Hemachandra
Ph.D. (Coadvised with SeeKiong Ng)
Recipient of NUS Research Achievement Award and Teaching Fellowship Scheme
B.Sc. in Physics with Minors in Computer Science and Mathematics > Mahidol University International College
Research Interests: deep active learning


Lin, Xiaoqiang
Ph.D. (Coadvised with SeeKiong Ng)
Recipient of NUS Research Achievement Award and Teaching Fellowship Scheme
B.Sc. in Statistics (Data Science & Technology) > Fudan University
Research Interests: incentives in collaborative machine learning and federated learning, data valuation


He, Zhenfeng
Ph.D.
B.Sc. in Computer Science (Hons. 1st Class) > National University of Singapore
Research Interests: automated machine learning


Sng, Weicong
Ph.D.
Recipient of Graduate Tutorship
B.Sc. in Statistics (Hons. 2nd Upper) > National University of Singapore
M.Comp. in Computer Science > National University of Singapore
Research Interests: reinforcement learning


Qiao, Rui
Ph.D.
Recipient of NUS Research Achievement Award, AI Singapore Ph.D. Fellowship, and Keppel Award of Excellence for the Top 2 students in B.Eng. in Information Systems Technology and Design (ISTD) pillar in Singapore University of Technology and Design
B.Eng. in Information Systems Technology and Design > Singapore University of Technology and Design
Research Interests: causal inference


Zhang, Zeyu
Ph.D. (Coadvised with Bolin Ding, Alibaba Group)
Recipient of Economic Development Board Industrial Postgraduate Programme (EDB IPP)
B.Eng. in Electrical and Electronic Engineering (Hons. 1st Class) with Minor in Physics and Mathematics > Nanyang Technological University
Research Interests: machine unlearning, federated learning


Wu, Zhaoxuan
Ph.D.
Recipient of Singapore Data Science Consortium (SDSC) Dissertation Research Fellowship, NUS Graduate School for Integrative Sciences and Engineering Scholarship (NGSS), Lijen Industrial Development Medal for the best student in the Honours Year term project in B.Sc. (Honours)  Data Science and Analytics programme
B.Sc. in Data Science and Analytics (Hons. 1st Class) > National University of Singapore
Research Interests: incentives in collaborative machine learning and federated learning, data valuation


Tay, Sebastian Shenghong
Ph.D. (Coadvised with ChuanSheng Foo, A*STAR)
Recipient of NUS Research Achievement Award x 2, A*STAR Computing and Information Science Scholarship (ACIS)
B.Sc. in Computer Science (Hons. 1st Class) > National University of Singapore
Research Interests: incentives in collaborative machine learning and federated learning, data valuation, Bayesian optimization


Xu, Xinyi 徐信羿
Ph.D. (Coadvised with ChuanSheng Foo, A*STAR)
Recipient of NUS Research Achievement Award x 2, A*STAR Computing and Information Science Scholarship (ACIS), Honor List of Student Tutors for Excellence in Teaching, Teaching Fellowship Scheme
B.Sc. in Computer Science (Hons. 1st Class) > National University of Singapore
Research Interests: incentives in collaborative machine learning and federated learning, reinforcement learning, data valuation


Sim, Rachael Hwee Ling
Ph.D. (Coadvised by Patrick Jaillet, MIT)
Recipient of NUS Research Achievement Award x 2, Teaching Fellowship Scheme (selected), SMART Graduate Fellowship, Lee Kuan Yew Gold Medal for best performing graduate in B.Comp. (Computer Science) programme, Tata Consultancy Services Asia Pacific Prize for Best Year 3 B.Comp. (Computer Science) student
B.Sc. in Computer Science (Hons. 1st Class) > National University of Singapore
Research Interests: incentives in collaborative machine learning and federated learning, data valuation, Bayesian optimization


Mohit Rajpal
Ph.D.
Recipient of NUS President's Graduate Fellowship
M.Sc. in Computer Science > Columbia University
B.Sc. in Computer Science > University of Illinois UrbanaChampaign
Research Interests: deep learning


Lucas Agussurja
Ph.D.
Recipient of NUS Research Achievement Award
B.Sc. in Computer Science (Hons. 2nd Upper Class) > National University of Singapore
Research Interests: data valuation


Arun Verma
Postdoctoral Fellow
Recipient of Naik and Rastogi Excellence in Ph.D. Thesis Award in IIT Bombay, COMSNETS 2022 Best Ph.D. Thesis Award, & Academic Excellence Award (Gold Medal) for highest CGPA in B.Tech. in CS & Eng., Shobhit University
B.Tech. in Computer Science and Engineering > Shobhit University, 2014
Ph.D. in Industrial Engineering and Operations Research > Indian Institute of Technology Bombay, Dec 2020
Ph.D. Thesis: Sequential Decision Problems with Weak Feedback
Research Interests: reinforcement learning, deep learning, and federated learning


Trần, Gia Lạc
Postdoctoral Fellow
B.Sc. in Computer Science > Honors Program of Faculty of Information Technology > Vietnam National University Ho Chi Minh City  University of Science, Sep 2014
M.Eng. in Data Science > Télécom Paris, Sep 2017
Ph.D. in Computer Science > Sorbonne Université, Dec 2020
Ph.D. Thesis: Advances of Deep Gaussian Processes: Calibration and Sparsification
Research Interests: Gaussian process

Former Members

Dai, Zhongxiang 代忠祥
Postdoctoral Fellow, Massachusetts Institute of Technology, Jan 2024 (Advised by Patrick Jaillet, MIT)
Recipient of NUS Dean's Graduate Research Excellence Award, NUS Research Achievement Award x 2, SMART Graduate Fellowship, & ST Electronics Prizes for being the top Year 1 and Year 2 student in Electrical Engineering
B.Eng. in Electrical Engineering (Hons. 1st Class) > National University of Singapore
Ph.D. in Computer Science > National University of Singapore (Coadvised by Patrick Jaillet, MIT), Jul 2021
Postdoctoral Fellow > National University of Singapore, Aug 2021  Dec 2023
Ph.D. Thesis: SampleEfficient Automated Machine Learning with Bayesian Optimization
Research Interests: Bayesian optimization


Shu, Yao 舒瑶
Senior Researcher, Tencent AI Lab, Jun 2023
Recipient of IMDA Excellence in Computing Prize (Best Ph.D. Thesis in NUS School of Computing) 2023 & NUS Dean's Graduate Research Excellence Award, Valedictorian for the class of NUS School of Computing Ph.D. graduates 2023
B.Sc. in Computer Science > Huazhong University of Science and Technology
Ph.D. in Computer Science > National University of Singapore, May 2022
Postdoctoral Fellow > National University of Singapore, Jun 2022  May 2023
Ph.D. Thesis: Understanding and Improving Neural Architecture Search
Research Interests: automated machine learning (AutoML), neural architecture search, hyperparameter optimization


Nguyễn, Quốc Phong
Postdoctoral Associate, Massachusetts Institute of Technology, Apr 2023 (Advised by Patrick Jaillet, MIT)
Recipient of NUS Research Achievement Award, SMART SMA3 Graduate Fellowship, Lee Kuan Yew Gold Medal for best performing graduate in B.Eng. (Computing Engineering) programme, & IES Gold Medal for top graduating student in B.Eng. in Computer Engineering
B.Eng. in Computer Engineering > National University of Singapore, 2013
Ph.D. in Computer Science > National University of Singapore (Coadvised by Patrick Jaillet, MIT), Dec 2018
Postdoctoral Fellow & Senior Postdoctoral Fellow > National University of Singapore, Jan 2019  Mar 2023
Ph.D. Thesis: An Alternative InformationTheoretic Criterion for Active Learning
Research Interests: probabilistic machine learning, Bayesian optimization, inverse reinforcement learning, collaborative machine learning
+ oral defense ◊ nov 13 ◊ 2K+18


Lam, Chi Thanh
Quantitative Researcher, Citadel Securities, Jun 2023
Recipient of NUS Research Achievement Award x 2, Honor List of Student Tutors for Excellence in Teaching, & SMART Graduate Fellowship
B.Sc. in Computer Science (Hons. 1st Class) > National University of Singapore
Ph.D. in Computer Science > National University of Singapore (Coadvised by Patrick Jaillet, MIT), May 2023
Ph.D. Thesis: Algorithms for Collaborative Machine Learning: Data Sharing and Model Sharing Perspectives
Research Interests: metalearning, collaborative machine learning, reinforcement learning


Sreejith Balakrishnan
Senior Responsible AI Scientist, Aicadium, Apr 2023
Recipient of NUS Research Achievement Award and Honor List of Student Tutors for Excellence in Teaching
B.Eng. in Electrical and Electronics Engineering (Hons. 1st Class) with Minor in Computing > Nanyang Technological University
M.Eng. in Electrical Engineering with Specialization in Computer Engineering > National University of Singapore
Ph.D. in Computer Science > National University of Singapore (Coadvised by Harold Soh), Mar 2023
Ph.D. Thesis: Towards HumanCentric AI: Inverse Reinforcement Learning Meets Algorithmic Fairness
Research Interests: robotics, machine learning


Chen, Yizhou
Machine Learning Engineer, Shopee, May 2022
B.Sc. in Physics (Hons. 1st Class) with Minor in Mathematics > Nanyang Technological University
Ph.D. in Computer Science > National University of Singapore, May 2022
Ph.D. Thesis: Exploiting Gradient Information for Modern Machine Learning Problems
Research Interests: metalearning, adversarial machine learning, deep Gaussian process


Teng, Tong 滕茼
Research Engineer, Huawei Technologies, Dec 2021
Recipient of NUS Research Achievement Award
B.Eng. in Computer Science > Shandong University
Ph.D. in Computer Science > National University of Singapore, Dec 2021
Ph.D. Thesis: Automated Kernel Selection for Gaussian Processes on Large Datasets
Research Interests: probabilistic machine learning


Dmitrii Kharkovskii Дмитрий Алексеевич Харьковский
Assistant Vice President, Data Science Lead, OCBC Group Data Office, Aug 2020
Recipient of NUS Research Achievement Award
Specialist degree in Math > St. Petersburg State University
Ph.D. in Computer Science > National University of Singapore, Dec 2020
Ph.D. Thesis: Automated Machine Learning: New Advances on Bayesian Optimization
Research Interests: privacypreserving machine learning, Bayesian optimization


Yu, Haibin 于海斌
Senior Research Engineer, Tencent, Shenzhen, Jan 2021
Recipient of NUS Research Achievement Award and SMART Graduate Fellowship
B.Eng. in Mechanical Engineering and Automation > Beihang University
Ph.D. in Computer Science > National University of Singapore, May 2020 (Coadvised by Patrick Jaillet, MIT)
Ph.D. Thesis: New Advances in Bayesian Inference for Gaussian Process and Deep Gaussian Process Models
Research Interests: Bayesian deep learning


Zhang, Yehong 张叶红
Research Scientist, Peng Cheng Laboratory, Shenzhen, Aug 2020
Recipient of AAAI 2016 Scholarship & NUS Research Achievement Award
B.Eng. in Computer Science > Harbin Institute of Technology
Ph.D. in Computer Science > National University of Singapore, Dec 2017 (Coadvised by Mohan Kankanhalli)
Ph.D. Thesis: DataEfficient Machine Learning with Multiple Output Types and High Input Dimensions
Research Interests: probabilistic machine learning, active learning, Bayesian optimization
+ oral defense ◊ nov 20 ◊ 2K+17


Chen, Jie 陈杰
Assistant Professor, College of Computer Science and Software Engineering, Shenzhen University, Apr 2018
Recipient of NUS Dean's Graduate Research Excellence Award, NUS Research Achievement Award, & UAI 2012 Scholarship
B.Eng. in Electrical Engineering > Taizhou University
M.Eng. in Computer Science > Zhejiang University
Ph.D. in Computer Science > National University of Singapore, Dec 2013
Postdoctoral Associate > SMART FM, Jan 2014  Mar 2018 (Coadvised by Patrick Jaillet, MIT)
Ph.D. Thesis: Gaussian ProcessBased Decentralized Data Fusion & Active Sensing Agents: Towards LargeScale Modeling & Prediction of Spatiotemporal Traffic Phenomena
Research Interests: robotics, multiagent planning, machine learning


Hoàng, Trọng Nghĩa
Assistant Professor, Washington State University, Jan 2023
Recipient of NUS Dean's Graduate Research Excellence Award, President's Graduate Fellowship, NUS Research Achievement Awards x 2, IJCAI 2013 Travel Grant Award, & AAMAS 2012 Scholarship
B.Sc. in Computer Science (Hons) > University of Science > Vietnam National University
Ph.D. in Computer Science > National University of Singapore, Feb 2015
Postdoctoral Fellow > Massachusetts Institute of Technology, Apr 2017 (Advised by Jonathan How)
Research Staff Member, MITIBM Watson AI Lab, Aug 2018
Senior Research Scientist, Amazon AWS AI Labs, Nov 2020
Ph.D. Thesis: New Advances on Bayesian and DecisionTheoretic Approaches for Interactive Machine Learning
Research Interests: machine learning, multiagent planning
+ oral defense ◊ jan 21 ◊ 2K+15


Ouyang, Ruofei 欧阳若飞
Senior Data Scientist, Shopee
Recipient of AAMAS 2014 Scholarship
B.Sc. in Computer Science > East China Normal University
Ph.D. in Computer Science > National University of Singapore, Dec 2016
Ph.D. Thesis: Exploiting Decentralized Multiagent Coordination for LargeScale Machine Learning Problems
Research Interests: Gaussian process, active sensing/learning, data fusion, Bayesian optimization
+ oral defense ◊ nov 7 ◊ 2K+16


Xu, Nuo 许诺
Backend Engineer, Grab
Recipient of AAAI 2014 Scholarship & NUS Research Achievement Award
Bachelor of Software Engineering > Harbin Institute of Technology
Ph.D. in Computer Science > National University of Singapore, Jan 2017
Ph.D. Thesis: Online Gaussian Process Filtering for Persistent Robot Localization With Arbitrary Sensor Modalities
Research Interests: machine learning, robotics
+ oral defense ◊ jan 11 ◊ 2K+17


Prabhu Natarajan
Lecturer, Department of Computer Science, National University of Singapore, May 2021
Recipient of ICDSC 2012 Best PhD Forum Paper Award, NUS Research Achievement Award, & AAMAS 2012 Scholarship
B.Tech. in Information Technology > Anna University
M.Eng. in Computer Science & Engineering > Anna University
Ph.D. in Computer Science > National University of Singapore, Dec 2013 (Coadvised by Mohan Kankanhalli)
Assistant Professor, DigiPen Institute of Technology Singapore, Jun 2016  Apr 2021
Ph.D. Thesis: A DecisionTheoretic Approach for Controlling & Coordinating Multiple Active Cameras in Surveillance
Research Interests: multicamera surveillance, decisiontheoretic planning & control for sensor networks
+ commencement ceremony ◊ aug 13 ◊ 2K+14


Lim, Kar Wai
Research Staff Member, IBM Research, Singapore
Recipient of ACML 2016 Best Student Paper Award & AMP Prize for Honours Thesis in Actuarial Studies (Best Thesis Award)
Bachelor of Actuarial Studies (Hons. 1st Class) > Australian National University
Ph.D. in Computer Science > Australian National University, Dec 2016
Research Fellow > NUSSingtel Cyber Security Research and Development Laboratory, Mar 2017  May 2019 (Coadvised by Mun Choon Chan)
Ph.D. Thesis: Nonparametric Bayesian Topic Modelling with Auxiliary Data
Research Interests: machine learning, Bayesian nonparametric methods, stochastic processes, point processes, Hawkes processes, Bayesian inference, Markov chain Monte Carlo methods


Tan, Wesley
Ph.D. in Computer Science, Nanyang Technological University, Aug 2017
Recipient of President's Graduate Fellowship in Nanyang Technological University
B.Sc. (Honors) in Economics > Purdue University
M.Sc. in Risk Management & Financial Engineering > Imperial College London, Oct 2015
M.Comp. in Computer Science > National University of Singapore, Jun 2017
M.Sc. Thesis: Variational Bayesian Actor Critic
Research Interests: reinforcement learning


Son, Jaemin 손재민
Bachelor of Comp Sci & Eng > Osaka University
M.Sc. in Computer Science > National University of Singapore, Nov 2016 (Coadvised by Gary Tan)
M.Sc. Thesis: HighDimensional Bayesian Optimization with Application to Traffic Simulation
Research Interests: machine learning


Cao, Nannan 曹楠楠
Former Research Assistant
Bachelor of Software Engineering > East China Normal University
M.Sc. in Computer Science > National University of Singapore, Sep 2012
M.Sc. Thesis: InformationTheoretic MultiRobot Path Planning
Research Interests: machine learning, Gaussian process, environmental sensing
+ commencement ceremony ◊ sep 19 ◊ 2K+13


Etkin Barış Özgül
B.Sc. in Computer Science > Bilkent University
M.Sc. in Computer Science > National University of Singapore, Jan 2017
M.Sc. Thesis: Shuttleline Routing for MobilityonDemand Systems with Ridesharing
Research Interests: AI, multiagent systems


Hoang, Quang Minh
Ph.D. in Computer Science, CMU, Aug 2018
Recipient of Lee Kuan Yew Gold Medal for best performing graduate in B.Comp. (Computational Biology) programme, Outstanding Undergraduate Researcher Prize in National University of Singapore, & ICML 2015 Scholarship
B.Comp. in Computational Biology (Hons. 1st Class) > National University of Singapore, 2016
FYP Dissertation: A Probabilistic Approach for Protein Function Prediction with Hierarchical Structured Outputs
Research Interests: machine learning


Ling, Chun Kai
Ph.D. in Computer Science, CMU, Aug 2017
Recipient of Lee Kuan Yew Gold Medal for best performing graduate in B.Eng. (Computing Engineering) programme, IES Gold Medal for top graduating student in B.Eng. in Computing Engineering, Defence Science Technology Agency Gold Medal for best local final year student for the degree of B.Eng. (Computer Engineering), Micron Prize for being one of the top two local Year 2 Computer Engineering students, & AlcatelLucent Telecommunications Prize for best performance in a module in the area of Communications and Networks in BEng (EE) or BEng (CEG) examinations
B.Eng. in Computer Engineering > National University of Singapore, 2015
FYP Dissertation: Planning and Learning in Spatiotemporal Environmental Phenomena
Research Interests: planning under uncertainty, machine learning


Erik Alexander Daxberger
Ph.D., Department of Engineering, Univ. Cambridge, Jan 2019
Recipient of CambridgeTübingen Ph.D. fellowship in machine learning, LMU research award for excellent students for the Bachelor's thesis, and LMUexchange and PROSA scholarships for a student exchange program at NUS & ICML 2017 Travel Award
B.Sc. in Computer Science > LudwigMaximiliansUniversität, Munich, 2017
B.Sc. Thesis: Distributed Batch Bayesian Optimization
Research Interests: machine learning, AI


Khor, ShiJie
Software Engineer, Google Singapore
Recipient of Lee Kuan Yew Gold Medal for best performing graduate in B.Comp. (Computer Science) programme, IEEE Singapore Computer Society Book Prize for the best student in the Honours Year term project, and Tata Consultancy Services Asia Pacific Prize
B.Comp. in Computer Science (Honours Highest Distinction) > National University of Singapore, 2016
FYP Dissertation: Kernel Search for Gaussian Processes
Research Interests: machine learning


Nathan Azaria
Software Engineer, Facebook London
Recipient of National Computer Systems Medal And Prize for the top student in B.Comp. (Computer Science) programme
B.Comp. in Computer Science (Honours Highest Distinction) > National University of Singapore, 2016
FYP Dissertation: Stochastic Variational Inference on MultiOutput Gaussian Process
Research Interests: machine learning


Lim, Keng Kiat
Software Engineer, Facebook HQ
B.Comp. in Computer Science > National University of Singapore, 2016
FYP Dissertation: Learning with HighDimensional Data
Research Interests: machine learning


Akshay Viswanathan
Software Engineer, Visa Inc.
B.Eng. in Computer Engineering (Hons. 1st Class) > National University of Singapore, 2015
FYP Dissertation: Scaling up Machine Learning Techniques via Parallelization for Large Data
Research Interests: machine learning


Shailendra Khemka
Business Solutions: Software Engineer, Deutsche Bank AG  SG Branch
Recipient of Tata Consultancy Services Asia Pacific Medal and Prize for 2nd best graduate throughout the course of study for B.Comp, Defence Science & Technology Agency Prize for top UROP student in B.Comp., & Sung Kah Kay Memorial Prize Winner in NUS University Scholars Programme (USP)
University Scholars Programme & von Neumann Programme for B.Comp in Computer Science > National University of Singapore, 2013
FYP Dissertation: Autonomous Search for Victims in a Disaster Situation
Research Interests: multiagent planning


Yu, Jiangbo 余江波
KAI Square
B.Sc. in Computer Science > Peking University
Research Interests: statistical machine learning

publications
RESEARCH SPOTLIGHTS ^{}lay articles for light reading^{}
ACCEPTED PAPERS & PREPRINTS
Coauthors : My students and postdocs Former thesis advisors Collaborators

Fairness in federated learning.
Xiaoqiang Lin^{}, Xinyi Xu^{}, Zhaoxuan Wu^{}, Rachael Hwee Ling Sim^{}, SeeKiong Ng^{}, ChuanSheng Foo^{}, Patrick Jaillet^{}, Trong Nghia Hoang^{} & Bryan Kian Hsiang Low.
In L. M. Nguyen, T. N. Hoang, P.Y. Chen, editors, Federated Learning:
Theory and Practice, chapter 8, pages 143160, Academic Press, 2024.
Abstract. Federated learning (FL) enables a form of collaboration among multiple clients in jointly learning a machine learning (ML) model without centralizing their local datasets. Like in any collaboration, it is imperative to guarantee fairness so that the clients are willing to participate. For instance, it is unfair if one client benefits significantly more than others, or if some client benefits disproportionately to its contribution in the collaboration. Additionally, it is also unfair if the ML model makes biased predictions against certain groups of clients. This chapter discusses three specific notions of fairness by highlighting their motivations from realworld usecases, examining several specific definitions for each notion and lastly describing the corresponding algorithms proposed to achieve each notion of fairness. At the end, this chapter will also summarize the identified gaps in current research efforts into open problems.

Federated sequential decision making: Bayesian optimization, reinforcement learning, and beyond.
Zhongxiang Dai^{}, Flint Xiaofeng Fan^{}, Cheston Tan^{}, Trong Nghia Hoang^{}, Bryan Kian Hsiang Low & Patrick Jaillet^{}.
In L. M. Nguyen, T. N. Hoang, P.Y. Chen, editors, Federated Learning:
Theory and Practice, chapter 14, pages 257279, Academic Press, 2024.
Abstract. Federated learning (FL) in its classic form involves the collaborative training of supervised learning models (e.g., neural networks) among multiple agents/clients. However, in addition to supervised learning, many other machine learning tasks which are inherently sequential decisionmaking problems, such as Bayesian optimization (BO) and reinforcement learning (RL), also find important applications in the federated setting. For example, the crucial problem of hyperparameter tuning of neural networks in the federated setting calls for algorithms for federated BO; collaborative clinical treatment recommendation among multiple hospitals is a natural application for federated RL. However, the extension of these classic sequential decisionmaking algorithms into the federated setting is faced with immense challenges. Firstly, these algorithms (e.g., BO and RL) have to be adapted to satisfy the core principles of FL. For example, consistent with the requirement of FL, the raw data (e.g., the history of observations in BO and the trajectories in RL) of every agent can never be shared with the other agents. Next, it is challenging to preserve the rigorous theoretical guarantees of these classic sequential decisionmaking algorithms (e.g., the sublinear regret upper bound of classic BO algorithms and the sample complexity of classic policy gradient algorithms for RL) and at the same time consistently improve their empirical performances by leveraging the federation of multiple agents. In this regard, a number of recent works have tackled these challenges and hence introduced federated versions of classic sequential decisionmaking algorithms (e.g., federated BO and federated RL algorithms) which satisfy the core principles of FL and are both theoretically grounded and practically effective. In light of these recent advances, this chapter discusses federated sequential decisionmaking problems with a focus on recent representative works on federated BO and federated RL, and describes open problems and potential future directions in these areas.

Data valuation in federated learning.
Zhaoxuan Wu^{}, Xinyi Xu^{}, Rachael Hwee Ling Sim^{}, Yao Shu^{}, Xiaoqiang Lin^{}, Lucas Agussurja^{}, Zhongxiang Dai^{}, SeeKiong Ng^{}, ChuanSheng Foo^{}, Patrick Jaillet^{}, Trong Nghia Hoang^{} & Bryan Kian Hsiang Low.
In L. M. Nguyen, T. N. Hoang, P.Y. Chen, editors, Federated Learning:
Theory and Practice, chapter 15, pages 281296, Academic Press, 2024.
Abstract. Federated learning (FL) has become an increasingly popular solution paradigm for enabling collaborative machine learning (CML) in which multiple clients can collaboratively train a common model without sharing their private training data with others. However, broad adoption of FL in practice is still limited, as clients can often be reluctant to participate in such federated effort unless their contributions are accurately recognized and fairly compensated. Data valuation is thus extensively required to measure the relative contributions among clients. In this chapter, we review data valuation methods in the conventional supervised CML setting, followed by extensions to the FL paradigm. To better address the challenge that the private data from local clients cannot be made available to the server, we further discuss many specialized data valuation methods developed for both horizontal and vertical FL in detail. Overall, this chapter aims to provide a comprehensive suite of data valuation tools to empower FL practitioners in various practical scenarios.

Incentives in federated learning.
Rachael Hwee Ling Sim^{}, Sebastian Shenghong Tay^{}, Xinyi Xu^{}, Yehong Zhang^{}, Zhaoxuan Wu^{}, Xiaoqiang Lin^{}, SeeKiong Ng^{}, ChuanSheng Foo^{}, Patrick Jaillet^{}, Trong Nghia Hoang^{} & Bryan Kian Hsiang Low.
In L. M. Nguyen, T. N. Hoang, P.Y. Chen, editors, Federated Learning:
Theory and Practice, chapter 16, pages 299309, Academic Press, 2024.
Abstract. This chapter explores incentive schemes that encourage clients to participate in federated learning (FL) and contribute more valuable data. Such schemes are important to enable collaboration in competitive situations where clients need justifiable incentives to participate and benefit others with information acquired at significant costs and resources, such as collecting and processing data, computing and communicating model updates, risking the privacy of data via shared model updates. Incentivization addresses these concerns through three key components: (1) fair contribution evaluation of each client's data, (2) client selection to maximize the utility of the global model, and (3) reward allocation to clients. Intuitively, clients desire higher valued rewards which should at least outweigh their costs. These and other requirements will be formally described as incentives. The chapter will also discuss some recent solutions and open problems to achieve these incentives in various settings, which includes settings where the contribution evaluation is declared or measured while the rewards can be monetary or modelbased.

PINNACLE: PINN Adaptive ColLocation and Experimental points selection.
Gregory Kang Ruey Lau^{}, Apivich Hemachandra^{}, SeeKiong Ng^{} & Bryan Kian Hsiang Low.
In Proceedings of the 12th International Conference on Learning Representations (ICLR24), Vienna, Austria, May 7  11, 2024.
5% acceptance rate (spotlight)
Abstract. PhysicsInformed Neural Networks (PINNs), which incorporate PDEs as soft constraints, train with a composite loss function that contains multiple training point types: different types of collocation points chosen during training to enforce each PDE and initial/boundary conditions, and experimental points which are usually costly to obtain via experiments or simulations. Training PINNs using this loss function is challenging as it typically requires selecting large numbers of points of different types, each with different training dynamics. Unlike past works that focused on the selection of either collocation or experimental points, this work introduces PINN Adaptive ColLocation and Experimental points selection (PINNACLE), the first algorithm that jointly optimizes the selection of all training point types, while automatically adjusting the proportion of collocation point types as training progresses. PINNACLE uses information on the interactions among training point types, which had not been considered before, based on an analysis of PINN training dynamics via the Neural Tangent Kernel (NTK). We theoretically show that the criterion used by PINNACLE is related to the PINN generalization error, and empirically demonstrate that PINNACLE is able to outperform existing point selection methods for forward, inverse, and transfer learning problems.

Robustifying and Boosting TrainingFree Neural Architecture Search.
Zhenfeng He^{}, Yao Shu^{}, Zhongxiang Dai^{} & Bryan Kian Hsiang Low.
In Proceedings of the 12th International Conference on Learning Representations (ICLR24), Vienna, Austria, May 7  11, 2024.
31% acceptance rate
Abstract. Neural architecture search (NAS) has become a key component of AutoML and a standard tool to automate the design of deep neural networks. Recently, trainingfree NAS as an emerging paradigm has successfully reduced the search costs of standard trainingbased NAS by estimating the true architecture performance with only trainingfree metrics. Nevertheless, the estimation ability of these metrics typically varies across different tasks, making it challenging to achieve robust and consistently good search performance on diverse tasks with only a single trainingfree metric. Meanwhile, the estimation gap between trainingfree metrics and the true architecture performances limits trainingfree NAS to achieve superior performance. To address these challenges, we propose the robustifying and boosting trainingfree NAS (RoBoT) algorithm which (a) employs the optimized combination of existing trainingfree metrics explored from Bayesian optimization to develop a robust and consistently betterperforming metric on diverse tasks, and (b) applies greedy search, i.e., the exploitation, on the newly developed metric to bridge the aforementioned gap and consequently to boost the search performance of standard trainingfree NAS further. Remarkably, the expected performance of our RoBoT can be theoretically guaranteed, which improves over the existing trainingfree NAS under mild conditions with additional interesting insights. Our extensive experiments on various NAS benchmark tasks yield substantial empirical evidence to support our theoretical results.

IncentiveAware Federated Learning with TrainingTime Model Rewards.
Zhaoxuan Wu^{}, Mohammad Mohammadi Amiri, Ramesh Raskar & Bryan Kian Hsiang Low.
In Proceedings of the 12th International Conference on Learning Representations (ICLR24), Vienna, Austria, May 7  11, 2024.
31% acceptance rate
Abstract. In federated learning (FL), incentivizing contributions of training resources (e.g., data, compute) from potentially competitive clients is crucial. Existing incentive mechanisms often distribute posttraining monetary rewards, which suffer from practical challenges of timeliness and feasibility of the rewards. Rewarding the clients after the completion of training may incentivize them to abort the collaboration, and monetizing the contribution is challenging in practice. To address these problems, we propose an incentiveaware algorithm that offers differentiated trainingtime model rewards for each client at each FL iteration. We theoretically prove that such a local design ensures the global objective of client incentivization. Through theoretical analyses, we further identify the issue of error propagation in model rewards and thus propose a stochastic referencemodel recovery strategy to ensure theoretically that all the clients eventually obtain the optimal model in the limit. We perform extensive experiments to demonstrate the superior incentivizing performance of our method compared to existing baselines.

Understanding Domain Generalization: A Noise Robustness Perspective.
Rui Qiao^{} &
Bryan Kian Hsiang Low.
In Proceedings of the 12th International Conference on Learning Representations (ICLR24), Vienna, Austria, May 7  11, 2024.
31% acceptance rate
Abstract. Despite the rapid development of machine learning algorithms for domain generalization (DG), there is no clear empirical evidence that the existing DG algorithms outperform the classic empirical risk minimization (ERM) across standard benchmarks. To better understand this phenomenon, we investigate whether there are benefits of DG algorithms over ERM through the lens of label noise. Specifically, our finitesample analysis reveals that label noise exacerbates the effect of spurious correlations for ERM, undermining generalization. Conversely, we illustrate that DG algorithms exhibit implicit labelnoise robustness during finitesample training even when spurious correlation is present. Such desirable property helps mitigate spurious correlations and improve generalization in synthetic experiments. However, additional comprehensive experiments on realworld benchmark datasets indicate that labelnoise robustness does not necessarily translate to better performance compared to ERM. We conjecture that the failure mode of ERM arising from spurious correlations may be less prevalent in practice.

A Unified Framework for Bayesian Optimization under Contextual Uncertainty.
Sebastian Shenghong Tay^{}, ChuanSheng Foo^{}, Daisuke Urano^{}, Richalynn Leong^{} &
Bryan Kian Hsiang Low.
In Proceedings of the 12th International Conference on Learning Representations (ICLR24), Vienna, Austria, May 7  11, 2024.
31% acceptance rate
Abstract. Bayesian optimization under contextual uncertainty (BOCU) is a family of BO problems in which the learner makes a decision prior to observing the context and must manage the risks involved. Distributionally robust BO (DRBO) is a subset of BOCU that affords robustness against context distribution shift, and includes the optimization of expected values and worstcase values as special cases. By considering the first derivatives of the DRBO objective, we generalize DRBO to one that includes several other uncertainty objectives studied in the BOCU literature such as worstcase sensitivity (and thus notions of risk such as variance, range, and conditional valueatrisk) and meanrisk tradeoffs. We develop a general Thompson sampling algorithm that is able to optimize any objective within the BOCU framework, analyze its theoretical properties, and compare it to suitable baselines across different experimental settings and uncertainty objectives.

Optimistic Bayesian Optimization with Unknown Constraints.
Quoc Phong Nguyen^{}, Wan Theng Ruth Chew^{}, Le Song^{},
Bryan Kian Hsiang Low & Patrick Jaillet^{}.
In Proceedings of the 12th International Conference on Learning Representations (ICLR24), Vienna, Austria, May 7  11, 2024.
31% acceptance rate
Abstract. Though some research efforts have been dedicated to constrained Bayesian optimization (BO), there remains a notable absence of a principled approach with a theoretical performance guarantee in the decoupled setting. Such a setting involves independent evaluations of the objective function and constraints at different inputs, and is hence a relaxation of the commonlystudied coupled setting where functions must be evaluated together. As a result, the decoupled setting requires an adaptive selection between evaluating either the objective function or a constraint, in addition to selecting an input (in the coupled setting). This paper presents a novel constrained BO algorithm with a provable performance guarantee that can address the above relaxed setting. Specifically, it considers the fundamental tradeoff between exploration and exploitation in constrained BO, and, interestingly, affords a noteworthy connection to active learning. The performance of our proposed algorithms is also empirically evaluated using several synthetic and realworld optimization problems.

Leveraging Previous Tasks in Optimizing Risk Measures with Gaussian Processes.
Quoc Phong Nguyen^{},
Bryan Kian Hsiang Low & Patrick Jaillet^{}.
In Proceedings of the 12th International Conference on Learning Representations (ICLR24), Vienna, Austria, May 7  11, 2024.
31% acceptance rate
Abstract. Research on optimizing the risk measure of a blackbox function using Gaussian processes, especially Bayesian optimization (BO) of risk measures, has become increasingly important due to the inevitable presence of uncontrollable variables in realworld applications. Nevertheless, existing works on BO of risk measures start the optimization from scratch for every new task without considering the results of previous tasks. In contrast, its vanilla BO counterpart has received a thorough investigation on utilizing previous tasks to speed up the current task through the body of works on metaBO which, however, have not considered risk measures. To bridge this gap, this paper presents the first algorithm for metaBO of risk measures (i.e., valueatrisk (VaR) and the conditional VaR) by introducing a novel adjustment to the upper confidence bound acquisition function. Our proposed algorithm exhibits two desirable properties: (i) invariance to scaling and vertical shifting of the blackbox function and (ii) robustness to previous harmful tasks. We provide a theoretical performance guarantee for our algorithm and empirically demonstrate its performance using several synthetic function benchmarks and realworld objective functions.
 WASA: WAtermarkbased Source Attribution for Large Language ModelGenerated Data.
Jingtan Wang^{}, Xinyang Lu^{}, Zitong Zhao^{}, Zhongxiang Dai^{}, ChuanSheng Foo^{}, SeeKiong Ng^{} & Bryan Kian Hsiang Low.
arXiv:2310.00646, Oct 1, 2023.
Abstract. The impressive performances of large language models (LLMs) and their immense potential for commercialization have given rise to serious concerns over the intellectual property (IP) of their training data.
In particular, the synthetic texts generated by LLMs may infringe the IP of the data being used to train the LLMs.
To this end, it is imperative to be able to (a) identify the data provider who contributed to the generation of a synthetic text by an LLM (source attribution) and (b) verify whether the text data from a data provider has been used to train an LLM (data provenance).
In this paper, we show that both problems can be solved by watermarking, i.e., by enabling an LLM to generate synthetic texts with embedded watermarks that contain information about their source(s).
We identify the key properties of such watermarking frameworks (e.g., source attribution accuracy, robustness against adversaries), and propose a WAtermarking for Source Attribution (WASA) framework that satisfies these key properties due to our algorithmic designs.
Our WASA framework enables an LLM to learn an accurate mapping from the texts of different data providers to their corresponding unique watermarks, which sets the foundation for effective source attribution (and hence data provenance).
Extensive empirical evaluations show that our WASA framework achieves effective source attribution and data provenance.
 Use Your INSTINCT: INSTruction optimization usIng Neural bandits Coupled with Transformers.
Xiaoqiang Lin^{}, Zhaoxuan Wu^{}, Zhongxiang Dai^{}, Wenyang Hu^{}, Yao Shu^{}, SeeKiong Ng^{}, Patrick Jaillet^{} & Bryan Kian Hsiang Low.
arXiv:2310.02905, Oct 2, 2023.
Abstract. Large language models (LLMs) have shown remarkable instructionfollowing capabilities and achieved impressive performances in various applications. However, the performances of LLMs depend heavily on the instructions given to them, which are typically manually tuned with substantial human efforts. Recent work has used the queryefficient Bayesian optimization (BO) algorithm to automatically optimize the instructions given to blackbox LLMs. However, BO usually falls short when optimizing highly sophisticated (e.g., highdimensional) objective functions, such as the functions mapping an instruction to the performance of an LLM. This is mainly due to the limited expressive power of the Gaussian process (GP) model which is used by BO as a surrogate to model the objective function. Meanwhile, it has been repeatedly shown that neural networks (NNs), especially pretrained transformers, possess strong expressive power and can model highly complex functions. So, we adopt a neural bandit algorithm which replaces the GP in BO by an NN surrogate to optimize instructions for blackbox LLMs. More importantly, the neural bandit algorithm allows us to naturally couple the NN surrogate with the hidden representation learned by a pretrained transformer (i.e., an opensource LLM), which significantly boosts its performance. These motivate us to propose our INSTruction optimization usIng Neural bandits Coupled with Transformers (INSTINCT) algorithm. We perform instruction optimization for ChatGPT and use extensive experiments to show that our INSTINCT consistently outperforms the existing methods in different tasks, such as in various instruction induction tasks and the task of improving the zeroshot chainofthought instruction.
 Goat: Finetuned LLaMA Outperforms GPT4 on Arithmetic Tasks.
Tiedong Liu^{} & Bryan Kian Hsiang Low.
arXiv:2305.14201, May 23, 2023.
Abstract. We introduce Goat, a finetuned LLaMA model that significantly outperforms GPT4 on a range of arithmetic tasks. Finetuned on a synthetically generated dataset, Goat achieves stateoftheart performance on BIGbench arithmetic subtask. In particular, the zeroshot Goat7B matches or even surpasses the accuracy achieved by the fewshot PaLM540B. Surprisingly, Goat can achieve nearperfect accuracy on largenumber addition and subtraction through supervised finetuning only, which is almost impossible with previous pretrained language models, such as Bloom, OPT, GPTNeoX, etc. We attribute Goat's exceptional performance to LLaMA's consistent tokenization of numbers. To tackle more challenging tasks like largenumber multiplication and division, we propose an approach that classifies tasks based on their learnability, and subsequently decomposes unlearnable tasks, such as multidigit multiplication and division, into a series of learnable tasks by leveraging basic arithmetic principles. We thoroughly examine the performance of our model, offering a comprehensive evaluation of the effectiveness of our proposed decomposition steps. Additionally, Goat7B can be easily trained using LoRA on a 24GB VRAM GPU, facilitating reproducibility for other researchers. We release our model, dataset, and the Python script for dataset generation.
 DeRDaVa: DeletionRobust Data Valuation for Machine Learning.
Xiao Tian^{}, Rachael Hwee Ling Sim^{}, Jue Fan^{} & Bryan Kian Hsiang Low.
In Proceedings of the 38th AAAI Conference on Artificial Intelligence (AAAI24), Vancouver, Canada, Feb 20  Feb 27, 2024.
23.75% acceptance rate
Abstract. Data valuation is concerned with determining a fair valuation of data from data sources to compensate them or to identify training examples that are the most or least useful for predictions. With the rising interest in personal data ownership and data protection regulations, model owners will likely have to fulfil more data deletion requests. This raises issues that have not been addressed by existing works: Are the data valuation scores still fair with deletions? Must the scores be expensively recomputed? The answer is no. To avoid recomputations, we propose using our data valuation framework DeRDaVa upfront for valuing each data source's contribution to preserving robust model performance after anticipated data deletions. DeRDaVa can be efficiently approximated and will assign higher value to data that are more useful or less likely to be deleted. We further generalize DeRDaVa to RiskDeRDaVa to cater to riskaverse/seeking model owners who are concerned with the worst/bestcases model utility. We also empirically demonstrate the practicality of our solutions.
 Incremental QuasiNewton Methods with Faster Superlinear Convergence Rates.
Zhuanghua Liu^{}, Luo Luo & Bryan Kian Hsiang Low.
In Proceedings of the 38th AAAI Conference on Artificial Intelligence (AAAI24), Vancouver, Canada, Feb 20  Feb 27, 2024.
23.75% acceptance rate
Abstract. We consider the finitesum optimization problem of the form min_{x ∈ Rd} f(x) = (1/n) (f_{1}(x)+ ... + f_{i}(x) + ... + f_{n}(x)) where each f_{i}(.) is strongly convex and has Lipschitz continuous gradient and Hessian. The recently proposed incremental quasiNewton method is based on BFGS update and achieves a local superlinear convergence rate of O((1 − (dκ)^{1})^{⌈t/n⌉2}) where κ is the condition number and t is the number of iterations. This paper proposes a more efficient quasiNewton method by incorporating the Symmetric Rank1 update into the incremental framework, which results in the better local convergence rate of O((1 − d^{1})^{⌈t/n⌉2}). Furthermore, we can boost our method by applying the block update on the Hessian approximation, which leads to the even faster convergence
rate of O((1 − (kd)^{1})^{⌈t/n⌉2}) where k < d is the rank of update matrix. The numerical experiments show the proposed methods significantly outperform the baseline methods.
 Decentralized SumofNonconvex Optimization.
Zhuanghua Liu^{} & Bryan Kian Hsiang Low.
In Proceedings of the 38th AAAI Conference on Artificial Intelligence (AAAI24), Vancouver, Canada, Feb 20  Feb 27, 2024.
23.75% acceptance rate
Abstract. We consider the optimization problem of minimizing the sumofnonconvex function, i.e., a convex function that is the average of nonconvex components. The existing stochastic algorithms for such a problem only focus on a single machine and the centralized scenario. In this paper, we study the sumofnonconvex optimization in the decentralized setting. We present a new theoretical analysis of the PMGTSVRG algorithm to this problem and prove the linear convergence of their approach. However, the convergence rate of the PMGTSVRG algorithm has a linear dependency on the condition number, which is undesirable for the illconditioned problem. To remedy this issue, we propose an accelerated stochastic decentralized firstorder algorithm by incorporating the techniques of acceleration, gradient tracking, and multiconsensus mixing into the SVRG algorithm.
The convergence rate of the proposed method has a squareroot dependency on the condition number. The numerical experiments validate the theoretical guarantee of our proposed algorithms on both synthetic and realworld datasets.
 Quantum Bayesian Optimization.
Zhongxiang Dai^{}, Gregory Kang Ruey Lau^{}, Arun Verma^{}, Yao Shu^{}, Bryan Kian Hsiang Low & Patrick Jaillet^{}.
In Advances in Neural Information Processing Systems 36: 37th Annual Conference on Neural Information Processing Systems (NeurIPS'23), New Orleans, LA, Dec 10  Dec 16, 2023.
26.1% acceptance rate
Abstract. Kernelized bandits, also known as Bayesian optimization (BO), has been a prevalent method for optimizing complicated blackbox reward functions. Various BO algorithms have been theoretically shown to enjoy upper bounds on their cumulative regret which are sublinear in the number T of iterations, and a regret lower bound of
Omega(sqrt(T)) has been derived which represents the unavoidable regrets for any classical BO algorithm. Recent works on quantum bandits have shown that with the aid of quantum computing, it is possible to achieve tighter regret upper bounds better than their corresponding classical lower bounds. However, these works are restricted to either multiarmed or linear bandits, and are hence not able to solve sophisticated realworld problems with nonlinear reward functions. To this end, we introduce the quantumGaussian processupper confidence bound (QGPUCB) algorithm. To the best of our knowledge, our QGPUCB is the first BO algorithm able to achieve a regret upper bound of O(poly log T), which is significantly smaller than its regret lower bound of
Omega(sqrt(T)) in the classical setting. Moreover, thanks to our novel analysis of the confidence ellipsoid, our QGPUCB with the linear kernel achieves a smaller regret than the quantum linear UCB algorithm from the previous work. We use simulations to verify that the theoretical quantum speedup achieved by our QGPUCB is also potentially relevant in practice.
 Incentives in Private Collaborative Machine Learning.
Rachael Hwee Ling Sim^{}, Yehong Zhang^{}, Trong Nghia Hoang^{}, Xinyi Xu^{}, Bryan Kian Hsiang Low & Patrick Jaillet^{}.
In Advances in Neural Information Processing Systems 36: 37th Annual Conference on Neural Information Processing Systems (NeurIPS'23), New Orleans, LA, Dec 10  Dec 16, 2023.
26.1% acceptance rate
Abstract. Collaborative machine learning involves training models on data from multiple parties but must incentivize their participation. Existing data valuation methods fairly value and reward each party based on shared data or model parameters but neglect the privacy risks involved. To address this, we introduce differential privacy (DP) as an incentive. Each party can select its required DP guarantee and perturb its sufficient statistic (SS) accordingly. The mediator values the perturbed SS by the Bayesian surprise it elicits about the model parameters. As our valuation function enforces a privacyvaluation tradeoff, parties are deterred from selecting excessive DP guarantees that reduce the utility of the grand coalition's model. Finally, the mediator rewards each party with different posterior samples of the model parameters. Such rewards still satisfy existing incentives like fairness but additionally preserve DP and a high similarity to the grand coalition's posterior. We empirically demonstrate the effectiveness and practicality of our approach on synthetic and realworld datasets.
 Model Shapley: Equitable Model Valuation with Blackbox Access.
Xinyi Xu^{}, Chi Thanh Lam^{}, ChuanSheng Foo^{} & Bryan Kian Hsiang Low.
In Advances in Neural Information Processing Systems 36: 37th Annual Conference on Neural Information Processing Systems (NeurIPS'23), New Orleans, LA, Dec 10  Dec 16, 2023.
26.1% acceptance rate
Abstract. Valuation methods of data and machine learning (ML) models are essential to the establishment of AI marketplaces. Importantly, certain practical considerations (e.g., operational constraints, legal restrictions) favor the use of model valuation over data valuation. Also, existing marketplaces that involve trading of pretrained ML models call for an equitable model valuation method to price them. In particular, we investigate the blackbox access setting which allows querying a model (to observe predictions) without disclosing modelspecific information (e.g., architecture and parameters). By exploiting a Dirichlet abstraction of a model's predictions, we propose a novel and equitable model valuation method called model Shapley. We also leverage a formal connection between the similarity in models and that in model Shapley values (MSVs) to devise a learning approach for predicting MSVs of many vendors' models (e.g., 150) in a largescale marketplace. We perform extensive empirical validation on the effectiveness of model Shapley using various realworld datasets and heterogeneous model types.

Bayesian Optimization with Costvarying Variable Subsets.
Sebastian Shenghong Tay^{}, ChuanSheng Foo^{}, Daisuke Urano^{}, Richalynn Leong^{} & Bryan Kian Hsiang Low.
In Advances in Neural Information Processing Systems 36: 37th Annual Conference on Neural Information Processing Systems (NeurIPS'23), New Orleans, LA, Dec 10  Dec 16, 2023.
26.1% acceptance rate
Abstract. We introduce the problem of Bayesian optimization with costvarying variable subsets (BOCVS) where in each iteration, the learner chooses a subset of query variables and specifies their values while the rest are randomly sampled. Each chosen subset has an associated cost. This presents the learner with the novel challenge of balancing between choosing more informative subsets for more directed learning versus leaving some variables to be randomly sampled to reduce incurred costs. This paper presents a novel Gaussian process upper confidence boundbased algorithm for solving the BOCVS problem that is provably noregret. We analyze how the availability of cheaper control sets helps in exploration and reduces overall regret. We empirically show that our proposed algorithm can find significantly better solutions than comparable baselines with the same budget.

Exploiting Correlated Auxiliary Feedback in Parameterized Bandits.
Arun Verma^{}, Zhongxiang Dai^{}, Yao Shu^{} & Bryan Kian Hsiang Low.
In Advances in Neural Information Processing Systems 36: 37th Annual Conference on Neural Information Processing Systems (NeurIPS'23), New Orleans, LA, Dec 10  Dec 16, 2023.
26.1% acceptance rate
Abstract. We study a novel variant of the parameterized bandits problem in which the learner can observe auxiliary feedback that is correlated with the observed reward. The auxiliary feedback is readily available in many reallife applications, e.g., an online platform that wants to recommend the bestrated services to its users can observe the user's rating of service (rewards) and collect additional information like service delivery time (auxiliary feedback). We first develop a method that exploits auxiliary feedback to build a reward estimator with tight confidence bounds, leading to a smaller regret. We then characterize the regret reduction in terms of the correlation coefficient between reward and auxiliary feedback. Experimental results in different settings also verify the performance gain achieved by our proposed method.
 Batch Bayesian Optimization For Replicable Experimental Design.
Zhongxiang Dai^{}, Quoc Phong Nguyen^{}, Sebastian Shenghong Tay^{}, Daisuke Urano^{}, Richalynn Leong^{}, Bryan Kian Hsiang Low & Patrick Jaillet^{}.
In Advances in Neural Information Processing Systems 36: 37th Annual Conference on Neural Information Processing Systems (NeurIPS'23), New Orleans, LA, Dec 10  Dec 16, 2023.
26.1% acceptance rate
Abstract. Many realworld experimental design problems (a) evaluate multiple experimental conditions in parallel and (b) replicate each condition multiple times due to large and heteroscedastic observation noise. Given a fixed total budget, this naturally induces a tradeoff between evaluating more unique conditions while replicating each of them fewer times vs. evaluating fewer unique conditions and replicating each more times. Moreover, in these problems, practitioners may be riskaverse and hence prefer an input with both good average performance and small variability. To tackle both challenges, we propose the Batch Thompson Sampling for Replicable Experimental Design (BTSRED) framework, which encompasses three algorithms. Our BTSREDKnown and BTSREDUnknown algorithms, for, respectively, known and unknown noise variance, choose the number of replications adaptively rather than deterministically such that an input with a larger noise variance is replicated more times. As a result, despite the noise heteroscedasticity, both algorithms enjoy a theoretical guarantee and are asymptotically noregret. Our MeanVarBTSRED algorithm aims at riskaverse optimization and is also asymptotically noregret. We also show the effectiveness of our algorithms in two practical realworld applications: precision agriculture and AutoML.

REFEREED PUBLICATIONS
dblp
Sorted by year 2K +
23 
22 
21 
20 
19 
18 
17 
16 
15 
14 
13 
12 
11 
10 
9 
8 
7 
6 
5 
4 
3 
2
 TrainingFree Neural Active Learning with InitializationRobustness Guarantees.
Apivich Hemachandra^{}, Zhongxiang Dai^{}, Jasraj Singh^{}, SeeKiong Ng^{} & Bryan Kian Hsiang Low.
In Proceedings of the 40th International Conference on Machine Learning (ICML23), pages 1293112971, Honolulu, HI, Jul 23  29, 2023.
27.9% acceptance rate
Abstract. Existing neural active learning algorithms have aimed to optimize the predictive performance of neural networks (NNs) by selecting data for labeling. However, other than a good predictive performance, being robust against random parameter initializations is also a crucial requirement in safetycritical applications. To this end, we introduce our expected variance with Gaussian processes (EVGP) criterion for neural active learning, which is theoretically guaranteed to select data points that lead to trained NNs with both good predictive performances and initialization robustness. Importantly, our EVGP criterion is trainingfree, i.e., it does not require any training of the NN during data selection, which makes it computationally efficient. We empirically demonstrate that our EVGP criterion is highly correlated with both initialization robustness and generalization performance, and show that it consistently outperforms baseline methods in terms of both desiderata, especially in situations with limited initial data or large batch sizes.
 Collaborative Causal Inference with Fair Incentives.
Rui Qiao^{}, Xinyi Xu^{} & Bryan Kian Hsiang Low.
In Proceedings of the 40th International Conference on Machine Learning (ICML23), pages 2830028320, Honolulu, HI, Jul 23  29, 2023.
27.9% acceptance rate
Abstract. Collaborative causal inference (CCI) aims to improve the estimation of the causal effect of treatment variables by utilizing data aggregated from multiple selfinterested parties. However, since their source data are valuable proprietary assets that can be costly or tedious to obtain, every party has to be incentivized to be willing to contribute to the collaboration, for example, with a guaranteed fair and sufficiently valuable reward (than performing causal inference on its own). This paper presents a reward scheme designed using the unique statistical properties required by causal inference to guarantee certain desirable incentive criteria (such as fairness and benefit) for the parties based on their contributions. To achieve this, we first propose a data valuation method to value parties' data for CCI based on the distributional closeness of its resulting treatment effect estimate to that utilizing the aggregated data from all parties. Then, we show how to value the parties' rewards fairly based on a modified variant of the Shapley value arising from our proposed data valuation for CCI. Finally, the Shapley fair rewards are realized in the form of improved and stochastically perturbed treatment effect estimates to be returned to the parties. We empirically demonstrate the effectiveness of our reward scheme using simulated and realworld datasets.
 Fair yet Asymptotically Equal Collaborative Learning.
Xiaoqiang Lin^{}, Xinyi Xu^{}, SeeKiong Ng^{}, ChuanSheng Foo^{} & Bryan Kian Hsiang Low.
In Proceedings of the 40th International Conference on Machine Learning (ICML23), pages 2122321259, Honolulu, HI, Jul 23  29, 2023.
27.9% acceptance rate
Abstract. In collaborative learning with streaming data, nodes (e.g., organizations) jointly and continuously learn a machine learning model by sharing the latest model updates computed from their latest streaming data. For the more resourceful nodes to be willing to share their model updates, they need to be fairly incentivized. This paper explores an incentive design that guarantees fairness so that nodes receive rewards commensurate to their contributions. Our approach leverages an explorethenexploit formulation to estimate the nodes' contributions (i.e., exploration) for realizing our theoretically guaranteed fair incentives (i.e., exploitation). However, we observe a "rich get richer" phenomenon arising from the existing approaches to guarantee fairness and it discourages the participation of the less resourceful nodes. To remedy this, we additionally preserve asymptotic equality, i.e., less resourceful nodes achieve equal performance eventually to the more resourceful/"rich" nodes. We empirically demonstrate in two settings with realworld streaming data: federated online incremental learning and federated reinforcement learning, that our proposed approach outperforms existing baselines in fairness and learning performance while remaining competitive in preserving equality.
 Pruning during Training by Network Efficacy Modeling.
Mohit Rajpal^{}, Yehong Zhang^{} & Kian Hsiang Low.
Machine Learning (Special Issue on ECMLPKDD 2022 Journal Track), volume 112, issue 7, pages 26532684, 2023.
Abstract. Deep neural networks (DNNs) are costly to train. Pruning, an approach to
alleviate model complexity by zeroing out or pruning DNN elements, has shown promise in reducing training costs for DNNs with little to no efficacy at a given task. This paper presents a novel method to perform early pruning of DNN elements (e.g., neurons or convolutional filters) during the training process while minimizing losses to model performance. To achieve this, we model the efficacy of DNN elements in a Bayesian manner conditioned upon efficacy data collected during the training and prune DNN elements with low predictive efficacy after training completion. Empirical evaluations show that the proposed Bayesian early pruning improves the computational efficiency of DNN training while better preserving model performance compared to other tested pruning approaches.
 FAIR: Fair Collaborative Active Learning with Individual Rationality for Scientific Discovery.
Xinyi Xu^{},
Zhaoxuan Wu^{}, Arun Verma^{}, ChuanSheng Foo^{} & Kian Hsiang Low.
In Proceedings of the 26th International Conference on Artificial Intelligence and Statistics (AISTATS23), pages 40334057, Valencia, Spain, Apr 25  27, 2023.
29% acceptance rate
Abstract. Scientific discovery aims to find new patterns and test specific hypotheses by analysing largescale experimental data. However, various practical limitations, like high experimental costs or the inability to do some experiments, make it challenging for researchers to collect sufficient experimental data for successful scientific discovery. To this end, we propose a framework named collaborative active learning (CAL) that enables researchers to share their experimental data for mutual benefit. Specifically, our proposed coordinated acquisition function sets out to achieve individual rationality and fairness so that everyone can equitably benefit from collaboration. Finally, we empirically demonstrate that our method outperforms existing batch active learning methods adapted to the CAL setting in terms of both learning performance and fairness on various realworld scientific discovery datasets (biochemistry, material science, and physics).
 NoRegret SampleEfficient Bayesian Optimization for Finding Nash Equilibria with Unknown Utilities.
Sebastian Tay^{},
Quoc Phong Nguyen^{}, ChuanSheng Foo^{} & Kian Hsiang Low.
In Proceedings of the 26th International Conference on Artificial Intelligence and Statistics (AISTATS23), pages 35913619, Valencia, Spain, Apr 25  27, 2023.
29% acceptance rate
Abstract. The Nash equilibrium (NE) is a classic solution concept for normalform games that is stable under potential unilateral deviations by selfinterested agents. Bayesian optimization (BO) has been used to find NE in continuous generalsum games with unknown costlytosample utility functions in a sampleefficient manner. This paper presents the first noregret BO algorithm that is sampleefficient in finding pure NE by leveraging theory on high probability confidence bounds with Gaussian processes and the maximum information gain of kernel functions. Unlike previous works, our algorithm is theoretically guaranteed to converge to the optimal solution (i.e., NE). We also introduce the novel setting of applying BO to finding mixed NE in unknown discrete generalsum games and show that our theoretical framework is general enough to be extended naturally to this setting by developing a noregret BO algorithm that is sampleefficient in finding mixed NE. We empirically show that our algorithms are competitive w.r.t. suitable baselines in finding NE.

RiskAware Reinforcement Learning with Coherent Risk Measures and NonLinear Function Approximation.
Chi Thanh Lam^{},
Arun Verma^{}, Kian Hsiang Low & Patrick Jaillet^{}.
In Proceedings of the 11th International Conference on Learning Representations (ICLR23), Kigali, Rwanda, May 1  5, 2023.
31.8% acceptance rate
Abstract. We study the riskaware reinforcement learning (RL) problem in the episodic finitehorizon Markov decision process with unknown transition and reward functions. In contrast to the riskneutral RL problem, we consider minimizing the risk of having low rewards, which arise due to the intrinsic randomness of the MDPs and imperfect knowledge of the model. Our work provides a unified framework to analyze the regret of riskaware RL policy with coherent risk measures in conjunction with nonlinear function approximation, which gives the first sublinear regret bounds in the setting. Finally, we validate our theoretical results via empirical experiments on synthetic and realworld data.

Federated Neural Bandits.
Zhongxiang Dai^{},
Yao Shu^{},
Arun Verma^{}, Flint Xiaofeng Fan^{}, Kian Hsiang Low & Patrick Jaillet^{}.
In Proceedings of the 11th International Conference on Learning Representations (ICLR23), Kigali, Rwanda, May 1  5, 2023.
31.8% acceptance rate
Abstract. Recent works on neural contextual bandits have achieved compelling performances due to their ability to leverage the strong representation power of neural networks (NNs) for reward prediction. Many applications of contextual bandits involve multiple agents who collaborate without sharing raw observations, thus giving rise to the setting of federated contextual bandits}. Existing works on federated contextual bandits rely on linear or kernelized bandits, which may fall short when modeling complex realworld reward functions. So, this paper introduces the federated neuralupper confidence bound (FNUCB) algorithm. To better exploit the federated setting, FNUCB adopts a weighted combination of two UCBs: UCB^{a} allows every agent to additionally use the observations from the other agents to accelerate exploration (without sharing raw observations), while UCB^{b} uses an NN with aggregated parameters for reward prediction in a similar way to federated averaging for supervised learning. Notably, the weight between the two UCBs required by our theoretical analysis is amenable to an interesting interpretation, which emphasizes UCB^{a} initially for accelerated exploration and relies more on UCB^{b} later after enough observations have been collected to train the NNs for accurate reward prediction (i.e., reliable exploitation). We prove sublinear upper bounds on both the cumulative regret and the number of communication rounds of FNUCB, and empirically demonstrate its competitive performance.
 ZerothOrder Optimization with TrajectoryInformed Derivative Estimation.
Yao Shu^{},
Zhongxiang Dai^{}, Weicong Sng^{}, Arun Verma^{}, Patrick Jaillet^{} & Kian Hsiang Low.
In Proceedings of the 11th International Conference on Learning Representations (ICLR23), Kigali, Rwanda, May 1  5, 2023.
31.8% acceptance rate
Abstract. Zerothorder (ZO) optimization, in which the derivative is unavailable, has recently succeeded in many important machine learning applications. Existing algorithms rely on finite difference (FD) methods for derivative estimation and gradient descent (GD)based approaches for optimization. However, these algorithms suffer from query inefficiency because additional function queries are required for derivative estimation in their every GD update, which typically hinders their deployment in applications where every function query is expensive. To this end, we propose a trajectoryinformed derivative estimation method which only uses the optimization trajectory (i.e., the history of function queries during optimization) and hence eliminates the need for additional function queries to estimate a derivative. Moreover, based on our derivative estimation, we propose the technique of dynamic virtual updates, which allows us to reliably perform multiple steps of GD updates without reapplying derivative estimation. Based on these two contributions, we introduce the zerothorder optimization with trajectoryinformed derivative estimation (ZoRD) algorithm for queryefficient ZO optimization. We theoretically demonstrate that our trajectoryinformed derivative estimation and our ZoRD algorithm improve over existing approaches, which is then supported by our realworld experiments such as blackbox adversarial attack, nondifferentiable metric optimization and derivativefree reinforcement learning.
 Recursive ReasoningBased TrainingTime Adversarial Machine Learning.
Yizhou Chen^{}, Zhongxiang Dai^{}, Haibin Yu^{}, Kian Hsiang Low & TeckHua Ho^{}.
Artificial Intelligence (Special Issue on RiskAware Autonomous Systems: Theory and Practice), volume 315, pages 103837, Feb 2023.
Abstract. The training process of a machine learning (ML) model may be subject to adversarial attacks from an attacker who attempts to undermine the test performance of the ML model by perturbing the training minibatches, and thus needs to be protected by a defender. Such a problem setting is referred to as trainingtime adversarial ML. We formulate it as a twoplayer game and propose a principled Recursive Reasoningbased TrainingTime adversarial ML (R2T2) framework to model this game. R2T2 models the reasoning process between the attacker and the defender and captures their bounded reasoning capabilities (due to bounded computational resources) through the recursive reasoning formalism. In particular, we associate a deeper level of recursive reasoning with the use of a higherorder gradient to derive the attack (defense) strategy, which naturally improves its performance while requiring greater computational resources. Interestingly, our R2T2 framework encompasses a variety of existing adversarial ML methods which correspond to attackers (defenders) with different recursive reasoning capabilities. We show how an R2T2 attacker (defender) can utilize our proposed nested projected gradient descentbased method to approximate the optimal attack (defense) strategy at an arbitrary level of reasoning. R2T2 can empirically achieve stateoftheart attack and defense performances on benchmark image datasets.
 Probably Approximate Shapley Fairness with Applications in Machine Learning.
Zijian Zhou^{},
Xinyi Xu^{}, Rachael Hwee Ling Sim^{}, ChuanSheng Foo^{} & Kian Hsiang Low.
In Proceedings of the 37th AAAI Conference on Artificial Intelligence (AAAI23), pages 59105918, Washington, DC, Feb 7  Feb 14, 2023.
19.6% acceptance rate (oral presentation)
Abstract. The Shapley value (SV) is adopted in various scenarios in machine learning including data valuation, agent valuation, feature attribution etc. because it guarantees fairness (i.e., satisfying several fairness axioms) and that fairness is important in such valuations (e.g., for pricing in a marketplace). However, the exact calculation of SVs (with exponential time complexity) is infeasible in practice, so estimations are necessary. Consequently, it raises the following important question: The exact SVs guarantee fairness, but do the SV estimates also guarantee the same fairness? We show the fairness guaranteed by exact SVs is too restrictive (for SV estimates) and generalise it to a probably approximate fairness. We propose fidelity score, a metric to measure the variation of SV estimates and establish theoretically the relationship between fidelity score and fairness guarantee. We exploit the relationship between fidelity score and fairness guarantee to propose a novel greedy active estimation (GAE) with theoretical guarantees. We theoretically show GAE achieves better fairness guarantee than the de facto MonteCarlo estimation and empirically verify GAE outperforms several existing methods in guaranteeing fairness while remaining competitive in estimation accuracy in various specific ML scenarios using realworld datasets.
 Tradeoff between Payoff and Model Rewards in ShapleyFair Collaborative Machine Learning.
Quoc Phong Nguyen^{}, Kian Hsiang Low & Patrick Jaillet^{}.
In Advances in Neural Information Processing Systems 35: 36th Annual Conference on Neural Information Processing Systems (NeurIPS'22), pages 3054230553, New Orleans, LA, Nov 28  Dec 3, 2022.
25.6% acceptance rate
Abstract. This paper investigates the problem of fairly trading off between payoff and model rewards in collaborative machine learning (ML) where parties aggregate their datasets together to obtain improved ML models over that of each party. Supposing parties can afford the optimal model trained on the aggregated dataset, we propose an allocation scheme that distributes the payoff fairly. Notably, the same scheme can be derived from two different approaches based on (1) desirable properties of the parties' payoffs or (2) that of the underlying payoff flows from one party to another. While the former is conceptually simpler, the latter can be used to handle the practical constraint on the budgets of parties. In particular, we propose desirable properties for achieving a fair adjustment of the payoff flows that can trade off between the model reward's performance and the payoff reward. We empirically demonstrate that our proposed scheme is a sensible solution in several scenarios of collaborative ML with different budget constraints.
 Unifying and Boosting GradientBased TrainingFree Neural Architecture Search.
Yao Shu^{}, Zhongxiang Dai^{}, Zhaoxuan Wu^{} & Kian Hsiang Low.
In Advances in Neural Information Processing Systems 35: 36th Annual Conference on Neural Information Processing Systems (NeurIPS'22), pages 3300133015, New Orleans, LA, Nov 28  Dec 3, 2022.
25.6% acceptance rate
Abstract. Neural architecture search (NAS) has gained immense popularity owing to its ability to automate neural architecture design. A number of trainingfree metrics are recently proposed to realize NAS without training, hence making NAS more scalable. Despite their competitive empirical performances, a unified theoretical understanding of these trainingfree metrics is lacking. As a consequence, (a) the relationships among these metrics are unclear, (b) there is no theoretical interpretation for their empirical performances, and (c) there may exist untapped potential in existing trainingfree NAS, which probably can be unveiled through a unified theoretical understanding. To this end, this paper presents a unified theoretical analysis of gradientbased trainingfree NAS, which allows us to (a) theoretically study their relationships, (b) theoretically guarantee their generalization performances, and (c) exploit our unified theoretical understanding to develop a novel framework named hybrid NAS (HNAS) which consistently boosts trainingfree NAS in a principled way. Remarkably, HNAS can enjoy the advantages of both trainingfree (i.e., superior search efficiency) and trainingbased (i.e., remarkable search effectiveness) NAS, which we have demonstrated through extensive experiments.
 SampleThenOptimize Batch Neural Thompson Sampling.
Zhongxiang Dai^{}, Yao Shu^{}, Kian Hsiang Low & Patrick Jaillet^{}.
In Advances in Neural Information Processing Systems 35: 36th Annual Conference on Neural Information Processing Systems (NeurIPS'22), pages 2333123344, New Orleans, LA, Nov 28  Dec 3, 2022.
25.6% acceptance rate
Abstract. Bayesian optimization (BO), which uses a Gaussian process (GP) as a surrogate to model its objective function, is popular for blackbox optimization. However, due to the limitations of GPs, BO underperforms in some problems such as those with categorical, highdimensional or image inputs. To this end, recent works have used the highly expressive neural networks (NNs) as the surrogate model and derived theoretical guarantees using the theory of neural tangent kernel (NTK). However, these works suffer from the limitations of the requirement to invert an extremely large parameter matrix and the restriction to the sequential (rather than batch) setting. To overcome these limitations, we introduce two algorithms based on the Thompson sampling (TS) policy named SampleThenOptimize Batch Neural TS (STOBNTS) and STOBNTSLinear. To choose an input query, we only need to train an NN (resp. a linear model) and then choose the query by maximizing the trained NN (resp. linear model), which is equivalently sampled from the GP posterior with the NTK as the kernel function. As a result, our algorithms sidestep the need to invert the large parameter matrix yet still preserve the validity of the TS policy. Next, we derive regret upper bounds for our algorithms with batch evaluations, and use insights from batch BO and NTK to show that they are asymptotically noregret under certain conditions. Finally, we verify their empirical effectiveness using practical AutoML and reinforcement learning experiments.
 On Provably Robust MetaBayesian Optimization.
Zhongxiang Dai^{}, Yizhou Chen^{}, Haibin Yu^{}, Kian Hsiang Low & Patrick Jaillet^{}.
In Proceedings of the 38th Conference on Uncertainty in Artificial Intelligence (UAI22), pages 475485, Eindhoven, Netherlands, Aug 15, 2022.
32.3% acceptance rate
Abstract. Bayesian optimization (BO) has become popular for sequential optimization of blackbox functions. When BO is used to optimize a target function, we often have access to previous evaluations of potentially related functions. This begs the question as to whether we can leverage these previous experiences to accelerate the current BO task through metalearning (metaBO), while ensuring robustness against potentially harmful dissimilar tasks that could sabotage the convergence of BO. This paper introduces two scalable and provably robust metaBO algorithms: robust metaGaussian processupper confidence bound (RMGPUCB) and RMGPThompson sampling (RMGPTS). We prove that both algorithms are asymptotically noregret even when some or all previous tasks are dissimilar to the current task, and show that RMGPUCB enjoys a better theoretical robustness than RMGPTS. We also exploit the theoretical guarantees to optimize the weights assigned to individual previous tasks through regret minimization via online learning, which diminishes the impact of dissimilar tasks and hence further enhances the robustness. Empirical evaluations show that (a) RMGPUCB performs effectively and consistently across various applications, and (b) RMGPTS, despite being less robust than RMGPUCB both in theory and in practice, performs competitively in some scenarios with less dissimilar tasks and is more computationally efficient.
 Neural Ensemble Search via Bayesian Sampling.
Yao Shu^{}, Yizhou Chen^{}, Zhongxiang Dai^{} & Kian Hsiang Low.
In Proceedings of the 38th Conference on Uncertainty in Artificial Intelligence (UAI22), pages 18031812, Eindhoven, Netherlands, Aug 15, 2022.
32.3% acceptance rate
Abstract. Recently, neural architecture search (NAS) has been applied to automate the design of neural networks in realworld applications. A large number of algorithms have been developed to improve the search cost or the performance of the final selected architectures in NAS. Unfortunately, these NAS algorithms aim to select only one single wellperforming architecture from their search spaces and thus have overlooked the capability of neural network ensemble (i.e., an ensemble of neural networks with diverse architectures) in achieving improved performance over a single final selected architecture. To this end, we introduce a novel neural ensemble search algorithm, called neural ensemble search via Bayesian sampling (NESBS), to effectively and efficiently select wellperforming neural network ensembles from a NAS search space. In our extensive experiments, NESBS algorithm is shown to be able to achieve improved performance over stateoftheart NAS algorithms while incurring a comparable search cost, indicating the superior performance of our NESBS algorithm over these conventional NAS algorithms in practice.
 On the Convergence of the Shapley Value in Parametric Bayesian Learning Games.
Lucas Agussurja^{}, Xinyi Xu^{} & Kian Hsiang Low.
In Proceedings of the 39th International Conference on Machine Learning (ICML22), pages 180196, Baltimore, MD, Jul 1723, 2022.
21.9% acceptance rate
Abstract. Measuring contributions is a classical problem in cooperative game theory where the Shapley value is the most wellknown solution concept. In this paper, we establish the convergence property of the Shapley value in parametric Bayesian learning games where players perform a Bayesian inference using their combined data, and the posteriorprior KL divergence is used as the characteristic function. We show that for any two players, under some regularity conditions, their difference in Shapley value converges in probability to the difference in Shapley value of a limiting game whose characteristic function is proportional to the logdeterminant of the joint Fisher information. As an application, we present an online collaborative learning framework that is asymptotically Shapleyfair. Our result enables this to be achieved without any costly computations of posteriorprior KL divergences. Only a consistent estimator of the Fisher information is needed. The framework's effectiveness is demonstrated with experiments using realworld data.
 DAVINZ: Data Valuation using Deep Neural Networks at Initialization.
Zhaoxuan Wu^{}, Yao Shu^{} & Kian Hsiang Low.
In Proceedings of the 39th International Conference on Machine Learning (ICML22), pages 2415024176, Baltimore, MD, Jul 1723, 2022.
21.9% acceptance rate
Abstract. Recent years have witnessed a surge of interest in developing trustworthy methods to evaluate the value of data in many realworld applications, e.g., collaborative machine learning, data marketplaces, etc. Existing data valuation methods typically valuate data using the generalization performance of converged machine learning models after their longterm model training, making data valuation on large complex deep neural networks (DNNs) unaffordable. To this end, we theoretically derive a domainaware generalization bound to estimate the generalization performance of DNNs without model training. We then exploit this theoretically derived generalization bound to develop a novel trainingfree data valuation method named data valuation at initialization (DAVINZ) on DNNs, which consistently achieves remarkable effectiveness and efficiency in practice. Moreover, our trainingfree DAVINZ, surprisingly, can even theoretically and empirically enjoy the desirable properties that trainingbased data valuation methods usually attain, making it more trustworthy in practice.
 Efficient Distributionally Robust Bayesian Optimization with Worstcase Sensitivity.
Sebastian Tay^{}, ChuanSheng Foo^{}, Urano Daisuke^{}, Richalynn Leong^{} & Kian Hsiang Low.
In Proceedings of the 39th International Conference on Machine Learning (ICML22), pages 2118021204, Baltimore, MD, Jul 1723, 2022.
21.9% acceptance rate
Abstract. In distributionally robust Bayesian optimization (DRBO), an exact computation of the worstcase expected value requires solving an expensive convex optimization problem. We develop a fast approximation of the worstcase expected value based on the notion of worstcase sensitivity that caters to arbitrary convex distribution distances. We provide a regret bound for our novel DRBO algorithm with the fast approximation, and empirically show it is competitive with that using the exact worstcase expected value while incurring significantly less computation time. In order to guide the choice of distribution distance to be used with DRBO, we show that our approximation implicitly optimizes an objective close to an interpretable risksensitive value.
 Bayesian Optimization under Stochastic Delayed Feedback.
Arun Verma^{}, Zhongxiang Dai^{} & Kian Hsiang Low.
In Proceedings of the 39th International Conference on Machine Learning (ICML22), pages 2214522167, Baltimore, MD, Jul 1723, 2022.
21.9% acceptance rate
Abstract. Bayesian optimization (BO) is a widelyused sequential method for zerothorder optimization of complex and expensivetocompute blackbox functions. The existing BO methods assume that the function evaluation (feedback) is available to the learner immediately or after a fixed delay. Such assumptions may not be practical in many reallife problems like clinical trials, online recommendations, and hyperparameters tuning, where feedback is available after a random delay. To benefit from the experimental parallelization in these problems, the learner needs to start new function evaluations without waiting for delayed feedback. In this paper, we consider the BO under stochastic delayed feedback problem. We propose algorithms with sublinear regret guarantees that efficiently address the dilemma of selecting new function queries while waiting for randomly delayed feedback. Building on our results, we also make novel contributions to batch BO and contextual Gaussian process bandits. Our experiments on synthetic and reallife datasets verify the performance of proposed algorithms.
 Data Valuation in Machine Learning: "Ingredients", Strategies, and Open Challenges.
Rachael Hwee Ling Sim^{}, Xinyi Xu^{}
& Kian Hsiang Low.
In Proceedings of the 31st International Joint Conference on Artificial Intelligence (IJCAI22), pages 56075614, Vienna, Austria, Jul 2329, 2022.
18.2% acceptance rate
Abstract. Data valuation in machine learning (ML) is an emerging research area that studies the worth of data in ML. Data valuation is used in collaborative ML to determine a fair compensation for every data owner and in interpretable ML to identify the most responsible, noisy, or misleading training examples. This paper presents a comprehensive technical survey that provides a new formal study of data valuation in ML through its "ingredients" and the corresponding properties, grounds the discussion of common desiderata satisfied by existing data valuation strategies on our proposed ingredients, and identifies open research challenges for designing new ingredients, data valuation strategies, and cost reduction techniques.
 Markov Chain Monte CarloBased Machine Unlearning: Unlearning What Needs to be Forgotten.
Quoc Phong Nguyen^{}, Ryutaro Oikawa^{}, Dinil Mon Divakaran^{}, Mun Choon Chan^{} & Kian Hsiang Low.
In Proceedings of the 17th ACM ASIA Conference on Computer and Communications Security (ACM ASIACCS'22), pages 351363, Nagasaki, Japan, May 30  Jun 3, 2022.
18.4% acceptance rate
Abstract. As the use of machine learning (ML) models is becoming increasingly popular in many realworld applications, there are practical challenges that need to be addressed for model maintenance. One such challenge is to 'undo' the effect of a specific subset of dataset used for training a model. This specific subset may contain malicious or adversarial data injected by an attacker, which affects the model performance. Another reason may be the need for a service provider to remove data pertaining to a specific user to respect the user's privacy. In both cases, the problem is to 'unlearn' a specific subset of the training data from a trained model without incurring the costly procedure of retraining the whole model from scratch. Towards this goal, this paper presents a Markov chain Monte Carlobased machine unlearning (MCU) algorithm. MCU helps to effectively and efficiently unlearn a trained model from subsets of training dataset. Furthermore, we show that with MCU, we are able to explain the effect of a subset of a training dataset on the model prediction. Thus, MCU is useful for examining subsets of data to identify the adversarial data to be removed. Similarly, MCU can be used to erase the lineage of a user's personal data from trained ML models, thus upholding a user's "right to be forgotten". We empirically evaluate the performance of our proposed MCU algorithm on realworld phishing and diabetes datasets. Results show that MCU can achieve a desirable performance by efficiently removing the effect of a subset of training dataset and outperform an existing algorithm that utilizes the remaining dataset.
 NASI: Label and Dataagnostic Neural Architecture Search at Initialization.
Yao Shu^{}, Shaofeng Cai,
Zhongxiang Dai^{}, Beng Chin Ooi^{} & Kian Hsiang Low.
In Proceedings of the 10th International Conference on Learning Representations (ICLR22), Apr 25  29, 2022.
32.29% acceptance rate
Abstract. Recent years have witnessed a surging interest in Neural Architecture Search (NAS). Various algorithms have been proposed to improve the search efficiency and effectiveness of NAS, i.e., to reduce the search cost and improve the generalization performance of the selected architectures, respectively. However, the search efficiency of these algorithms is severely limited by the need for model training during the search process. To overcome this limitation, we propose a novel NAS algorithm called NAS at Initialization (NASI) that exploits the capability of a Neural Tangent Kernel in being able to characterize the performance of candidate architectures at initialization, hence allowing model training to be completely avoided to boost the search efficiency. Besides the improved search efficiency, NASI also achieves competitive search effectiveness on various datasets like CIFAR10/100 and ImageNet. Further, NASI is shown to be label and dataagnostic under mild conditions, which guarantees the transferability of architectures selected by our NASI over different datasets.
 NearOptimal Task Selection for MetaLearning with Mutual
Information and Online Variational Bayesian Unlearning.
Yizhou Chen^{},
Shizhuo Zhang^{} & Kian Hsiang Low.
In Proceedings of the 25th International Conference on Artificial Intelligence and Statistics (AISTATS22), pages 90919113, Mar 28  30, 2022.
29.2% acceptance rate
Abstract. This paper addresses the problem of active task selection which involves selecting the most informative tasks for metalearning. We propose a novel active task selection criterion based on the mutual information between latent task vectors. Unfortunately, such a criterion scales poorly in the number of candidate tasks when optimized. To resolve this issue, we exploit the submodularity property of our new criterion for devising the first active task selection algorithm for metalearning with a nearoptimal performance guarantee. To further improve the efficiency of our algorithm, we propose an online variant of the Stein variational gradient descent to perform fast belief updates of the metaparameters via maintaining a set of forward (and backward) particles when learning (or unlearning) from each selected task. We empirically demonstrate the performance of our proposed algorithm on realworld datasets.
 Incentivizing Collaboration in Machine Learning via Synthetic Data Rewards.
Sebastian Tay^{},
Xinyi Xu^{}, ChuanSheng Foo^{} & Kian Hsiang Low.
In Proceedings of the 36th AAAI Conference on Artificial Intelligence (AAAI22), pages 94489456, Feb 22  Mar 1, 2022.
4.26% acceptance rate (oral presentation)
Abstract. This paper presents a novel collaborative generative modeling (CGM) framework that incentivizes collaboration among selfinterested parties to contribute data to a pool for training a generative model (e.g., GAN), from which synthetic data are drawn and distributed to the parties as rewards commensurate to their contributions. Distributing synthetic data as rewards (instead of trained models or money) offers task and modelagnostic benefits for downstream learning tasks and is less likely to violate data privacy regulation. To realize the framework, we firstly propose a data valuation function using maximum mean discrepancy (MMD) that values data based on its quantity and quality in terms of its closeness to the true data distribution and provide theoretical results guiding the kernel choice in our MMDbased data valuation function. Then, we formulate the reward scheme as a linear optimization problem that when solved, guarantees certain incentives such as fairness in the CGM framework. We devise a weighted sampling algorithm for generating synthetic data to be distributed to each party as reward such that the value of its data and the synthetic data combined matches its assigned reward value by the reward scheme. We empirically show using simulated and realworld datasets that the parties' synthetic data rewards are commensurate to their contributions.
 Differentially Private Federated Bayesian Optimization with Distributed Exploration.
Zhongxiang Dai^{}, Kian Hsiang Low & Patrick Jaillet^{}.
In Advances in Neural Information Processing Systems 34: 35th Annual Conference on Neural Information Processing Systems (NeurIPS'21), pages 91259139, Dec 614, 2021.
25.6% acceptance rate
Abstract. Bayesian optimization (BO) has recently been extended to the federated learning (FL) setting by the federated Thompson sampling (FTS) algorithm. However, FTS is not equipped with a rigorous privacy guarantee which is an important consideration in FL. Recent works have incorporated differential privacy (DP) into the training of deep neural networks through a general framework for adding DP to iterative algorithms. Following this general DP framework, our work here integrates DP into FTS to preserve userlevel privacy. We also leverage the ability of this general DP framework to handle different parameter vectors, as well as the technique of local modeling for BO, to further improve the utility of our algorithm through distributed exploration (DE). The resulting differentially private FTS with DE (DPFTSDE) algorithm is endowed with theoretical guarantees for both the privacy and utility and is amenable to interesting theoretical insights about the privacyutility tradeoff. We also use realworld experiments to show that DPFTSDE achieves high utility (competitive performance) with a strong privacy guarantee (small privacy loss) and induces a practical tradeoff between privacy and utility.
 Optimizing Conditional ValueAtRisk of BlackBox Functions.
Quoc Phong Nguyen^{}, Zhongxiang Dai^{}, Kian Hsiang Low & Patrick Jaillet^{}.
In Advances in Neural Information Processing Systems 34: 35th Annual Conference on Neural Information Processing Systems (NeurIPS'21), pages 41704180, Dec 614, 2021.
25.6% acceptance rate
Abstract. This paper presents two Bayesian optimization (BO) algorithms with theoretical performance guarantee to maximize the conditional valueatrisk (CVaR) of a blackbox function: CVUCB and CVTS which are based on the wellestablished principle of optimism in the face of uncertainty and Thompson sampling, respectively. To achieve this, we develop an upper confidence bound of CVaR and prove the noregret guarantee of CVUCB by utilizing an interesting connection between CVaR and valueatrisk (VaR). For CVTS, though it is straightforwardly performed with Thompson sampling, bounding its Bayesian regret is nontrivial because it requires a tail expectation bound for the distribution of CVaR of a blackbox function, which has not been shown in the literature. The performances of both CVUCB and CVTS are empirically evaluated in optimizing CVaR of synthetic benchmark functions and simulated realworld optimization problems.
 FaultTolerant Federated Reinforcement Learning with Theoretical Guarantee.
Xiaofeng Fan^{}, Yining Ma, Zhongxiang Dai^{}, Wei Jing, Cheston Tan^{} & Kian Hsiang Low.
In Advances in Neural Information Processing Systems 34: 35th Annual Conference on Neural Information Processing Systems (NeurIPS'21), pages 10071021, Dec 614, 2021.
25.6% acceptance rate
Abstract. The growing literature of Federated Learning (FL) has recently inspired Federated Reinforcement Learning (FRL) to encourage multiple agents to federatively build a better decisionmaking policy without sharing raw trajectories. Despite its promising applications, existing FRL work fails to I) provide theoretical analysis on its convergence; II) account for random system failures and adversarial attacks. Towards this end, we propose the first FRL framework, the convergence of which is tolerant to less than half of participating agents being random system failures or adversarial attackers. We prove that the sample efficiency of the proposed framework is guaranteed to scale with the number of agents, accounting for such potential failures or attacks. We empirically verify all theoretical results on various RL benchmarking tasks.
 Gradient Driven Rewards to Guarantee Fairness in Collaborative Machine Learning.
Xinyi Xu^{}, Lingjuan Lyu, Xingjun Ma, Chenglin Miao, ChuanSheng Foo^{} & Kian Hsiang Low.
In Advances in Neural Information Processing Systems 34: 35th Annual Conference on Neural Information Processing Systems (NeurIPS'21), pages 1610416117, Dec 614, 2021.
25.6% acceptance rate
Abstract. In collaborative machine learning (CML), multiple agents pool their resources (e.g., data) together for a common learning task. In realistic CML settings where the agents are selfinterested and not altruistic, they may be unwilling to share data or model information without adequate rewards. Furthermore, as the data/model information shared by the agents may differ in quality, designing rewards which are fair to them is important so that they would not feel exploited nor discouraged from sharing. In this paper, we adopt federated learning as the CML paradigm, propose a novel cosine gradient Shapley value (CGSV) to fairly evaluate the expected marginal contribution of each agent's uploaded model parameter update/gradient without needing an auxiliary validation dataset, and based on the CGSV, design a novel trainingtime gradient reward mechanism with a fairness guarantee by sparsifying the aggregated parameter update/gradient downloaded from the server as reward to each agent such that its resulting quality is commensurate to that of the agent’s uploaded parameter update/gradient. We empirically demonstrate the effectiveness of our fair gradient reward mechanism on multiple benchmark datasets in terms of fairness, predictive performance, and time overhead.
 Validation Free and Replication Robust Volumebased Data Valuation.
Xinyi Xu^{}, Zhaoxuan Wu^{}, ChuanSheng Foo^{} & Kian Hsiang Low.
In Advances in Neural Information Processing Systems 34: 35th Annual Conference on Neural Information Processing Systems (NeurIPS'21), pages 1083710848, Dec 614, 2021.
25.6% acceptance rate
Abstract. Data valuation arises as a nontrivial challenge in use cases such as collaborative data sharing, data markets and etc. The value of data is often related to the learning performance, e.g. validation accuracy, of the model trained on the data. While intuitive, this methodology introduces a high coupling between data valuation and validation, which may not be desirable in practice. For instance, data providers may disagree on the choice of the validation set, or the validation set may be (statistically) different from the actual application. A separate but practical issue is data replication. If some data points are valuable, a dishonest data provider may offer a dataset containing replications of these data points, trying to exploit the valuation to get a higher reward/payment. Based on the ordinary least squares framework, our data valuation method does not require validation, and still provides a useful connection between the value of data and learning performance. In particular, we utilize the volume of the data matrix (determinant of its left Gram), thus able to provide an intuitive interpretation of the value of data via the diversity in the data. Furthermore, we formalize the robustness to data replication, and propose a robust volume valuation with robustness guarantees. We conduct extensive experiments to demonstrate its consistency and practical advantages over existing baselines.
 Learning to Learn with Gaussian Processes.
Quoc Phong Nguyen^{}, Kian Hsiang Low & Patrick Jaillet^{}.
In Proceedings of the 37th Conference on Uncertainty in Artificial Intelligence (UAI21), pages 14661475, Jul 2730, 2021.
26.5% acceptance rate
Abstract. This paper presents Gaussian process metalearning (GPML) for fewshot regression, which explicitly exploits the distance between regression problems/tasks using a novel task kernel. It contrasts sharply with the popular metricbased metalearning approach which is based on the distance between data inputs or their embeddings in the fewshot learning literature. Apart from the superior predictive performance by capturing the diversity of different tasks, GPML offers a set of representative tasks that are useful for understanding the task distribution. We empirically demonstrate the performance and interpretability of GPML in several fewshot regression problems involving a multimodal task distribution and realworld datasets.
 TrustedMaximizers Entropy Search for Efﬁcient Bayesian Optimization.
Quoc Phong Nguyen^{}, Zhaoxuan Wu^{}, Kian Hsiang Low & Patrick Jaillet^{}.
In Proceedings of the 37th Conference on Uncertainty in Artificial Intelligence (UAI21), pages 14861495, Jul 2730, 2021.
26.5% acceptance rate
Abstract. Informationbased Bayesian optimization (BO) algorithms have achieved stateoftheart performance in optimizing a blackbox objective function. However, they usually require several approximations or simplifying assumptions (without clearly understanding their effects on the BO performance) and/or their generalization to batch BO is computationally unwieldy, especially with an increasing batch size. To alleviate these issues, this paper presents a novel trustedmaximizers entropy search (TES) acquisition function: It measures how much an input query contributes to the information gain on the maximizer over a ﬁnite set of trusted maximizers, i.e., inputs optimizing functions that are sampled from the Gaussian process posterior belief of the objective function. Evaluating TES requires either only a stochastic approximation with sampling or a deterministic approximation with expectation propagation, both of which are investigated and empirically evaluated using synthetic benchmark objective functions and realworld optimization problems, e.g., hyperparameter tuning of a convolutional neural network and synthesizing ‘physically realizable’ faces to fool a blackbox face recognition system. Though TES can naturally be generalized to a batch variant with either approximation, the latter is amenable to be scaled to a much larger batch size in our experiments.
 Collaborative Bayesian Optimization with Fair Regret.
Rachael Hwee Ling Sim^{}, Yehong Zhang^{}, Kian Hsiang Low & Patrick Jaillet^{}.
In Proceedings of the 38th International Conference on Machine Learning (ICML21), pages 96919701, Jul 1824, 2021.
21.5% acceptance rate
Abstract. Bayesian optimization (BO) is a popular tool for optimizing complex and costlytoevaluate blackbox objective functions. To further reduce the number of function evaluations, any party performing BO may be interested to collaborate with others to optimize the same objective function concurrently. To do this, existing BO algorithms have considered optimizing a batch of input queries in parallel and provided theoretical bounds on their cumulative regret reflecting inefficiency. However, when the objective function values are correlated with realworld rewards (e.g., money), parties may be hesitant to collaborate if they risk incurring larger cumulative regret (i.e., smaller realworld reward) than others. This paper shows that fairness and efficiency are both necessary for the collaborative BO setting. Inspired by social welfare concepts from economics, we propose a new notion of regret capturing these properties and a collaborative BO algorithm whose convergence rate can be theoretically guaranteed by bounding the new regret, both of which share an adjustable parameter for trading off between fairness vs. efficiency. We empirically demonstrate the benefits (e.g., increased fairness) of our algorithm using synthetic and realworld datasets.
 Model Fusion for Personalized Learning.
Chi Thanh Lam^{}, Trong Nghia Hoang^{}, Kian Hsiang Low & Patrick Jaillet^{}.
In Proceedings of the 38th International Conference on Machine Learning (ICML21), pages 59485958, Jul 1824, 2021.
21.5% acceptance rate
Abstract. Production systems operating on a growing domain of analytic services often require generating warmstart solution models for emerging tasks with limited data. One potential approach to address this challenge is to adopt meta learning to generate a base model that can be adapted to solve unseen tasks with minimal finetuning. This however requires the training processes of previous solution models of existing tasks to be synchronized. This is not possible if these models were pretrained separately on private data owned by different entities and cannot be synchronously retrained. To accommodate for such scenarios, we develop a new personalized learning framework that synthesizes customized models for unseen tasks via fusion of independently pretrained models of related tasks. We establish performance guarantee for the proposed framework and demonstrate its effectiveness on both synthetic and real datasets.
 ValueatRisk Optimization with Gaussian Processes.
Quoc Phong Nguyen^{}, Zhongxiang Dai^{}, Kian Hsiang Low & Patrick Jaillet^{}.
In Proceedings of the 38th International Conference on Machine Learning (ICML21), pages 80638072, Jul 1824, 2021.
21.5% acceptance rate
Abstract. Valueatrisk (VaR) is an established measure to assess risks in critical realworld applications with random environmental factors. This paper presents a novel VaR upper confidence bound (VUCB) algorithm for maximizing the VaR of a blackbox objective function with the first noregret guarantee. To realize this, we first derive a confidence bound of VaR and then prove the existence of values of the environmental random variable (to be selected to achieve no regret) such that the confidence bound of VaR lies within that of the objective function evaluated at such values. Our VUCB algorithm empirically demonstrates stateoftheart performance in optimizing synthetic benchmark functions, a portfolio optimization problem, and a simulated robot task.
 AID: Active Distillation Machine to Leverage PreTrained BlackBox Models in Private Data Settings.
Trong Nghia Hoang^{}, Shenda Hong, Cao Xiao, Kian Hsiang Low & Jimeng Sun.
In Proceedings of the 30th The Web Conference (WWW'21), pages 3569–3581, Apr 1923, 2021.
20.6% acceptance rate
Abstract. This paper presents an active distillation method for a local institution (e.g., hospital) to find the best queries within its given budget to distill an onserver blackbox model's predictive knowledge into a local surrogate with transparent parameterization. This allows local institutions to understand better the predictive reasoning of the blackbox model in its own local context or to further customize the distilled knowledge with its private dataset that cannot be centralized and fed into the server model. The proposed method thus addresses several challenges of deploying machine learning in many industrial settings (e.g., healthcare analytics) with strong proprietary constraints. These include: (1) the opaqueness of the server model’s architecture which prevents local users from understanding its predictive reasoning in their local data contexts; (2) the increasing cost and risk of uploading local data on the cloud for analysis; and (3) the need to customize the server model with private onsite data. We evaluated the proposed method on both benchmark and realworld healthcare data where significant improvements over existing local distillation methods were observed. A theoretical analysis of the proposed method is also presented.
 Topk Ranking Bayesian Optimization.
Quoc Phong Nguyen^{}, Sebastian Tay^{}, Kian Hsiang Low & Patrick Jaillet^{}.
In Proceedings of the 35th AAAI Conference on Artificial Intelligence (AAAI21), pages 91359143, Feb 29, 2021.
21.4% acceptance rate
Abstract. This paper presents a novel approach to topk ranking Bayesian optimization (topk ranking BO) which is a practical and significant generalization of preferential BO to handle topk ranking and tie/indifference observations. We first design a surrogate model that is not only capable of catering to the above observations, but is also supported by a classic random utility model. Another equally important contribution is the introduction of the first informationtheoretic acquisition function in BO with preferential observation called multinomial predictive entropy search (MPES) which is flexible in handling these observations and optimized for all inputs of a query jointly. MPES possesses superior performance compared with existing acquisition functions that select the inputs of a query one at a time greedily. We empirically evaluate the performance of MPES using several synthetic benchmark functions, CIFAR10 dataset, and SUSHI preference dataset.
 An InformationTheoretic Framework for Unifying Active Learning Problems.
Quoc Phong Nguyen^{}, Kian Hsiang Low & Patrick Jaillet^{}.
In Proceedings of the 35th AAAI Conference on Artificial Intelligence (AAAI21), pages 91269134, Feb 29, 2021.
21.4% acceptance rate
Abstract. This paper presents an informationtheoretic framework for unifying active learning problems: level set estimation (LSE), Bayesian optimization (BO), and their generalized variant. We first introduce a novel active learning criterion that subsumes an existing LSE algorithm and achieves stateoftheart performance in LSE problems with a continuous input domain. Then, by exploiting the relationship between LSE and BO, we design a competitive informationtheoretic acquisition function for BO that has interesting connections to upper confidence bound and maxvalue entropy search (MES). The latter connection reveals a drawback of MES which has important implications on not only MES but also on other MESbased acquisition functions. Finally, our unifying informationtheoretic framework can be applied to solve a generalized problem of LSE and BO involving multiple level sets in a dataefficient manner. We empirically evaluate the performance of our proposed algorithms using synthetic benchmark functions, a realworld dataset, and in hyperparameter tuning of machine learning models.
 Convolutional Normalizing Flows for Deep Gaussian Processes.
Haibin Yu^{}, Dapeng Liu, Kian Hsiang Low & Patrick Jaillet^{}.
In Proceedings of the International Joint Conference on Neural Networks (IJCNN'21), Jul 1822, 2021.
Abstract. Deep Gaussian processes (DGPs), a hierarchical composition of GP models, have successfully boosted the expressive power of their singlelayer counterpart. However, it is impossible to perform exact inference in DGPs, which has motivated the recent development of variational inferencebased methods. Unfortunately, either these methods yield a biased posterior belief or it is difficult to evaluate their convergence. This paper introduces a new approach for specifying flexible, arbitrarily complex, and scalable approximate posterior distributions. The posterior distribution is constructed through a normalizing flow (NF) which transforms a simple initial probability into a more complex one through a sequence of invertible transformations. Moreover, a novel convolutional normalizing flow (CNF) is developed to improve the time efficiency and capture dependency between layers. Empirical evaluation shows that CNF DGP outperforms the stateoftheart approximation methods for DGPs.
 Efficient Exploration of Reward Functions in Inverse Reinforcement Learning via Bayesian Optimization.
Sreejith Balakrishnan^{}, Quoc Phong Nguyen^{}, Kian Hsiang Low & Harold Soh^{}.
In Advances in Neural Information Processing Systems 33: 34th Annual Conference on Neural Information Processing Systems (NeurIPS'20), pages 41874198, Dec 612, 2020.
20.1% acceptance rate
Abstract. In this paper, we focus on the problem of Inverse Reinforcement Learning (IRL), which is relevant for a variety of tasks including value alignment and robot learning from demonstration. Despite significant algorithmic contributions in recent years, IRL remains an illposed problem at its core; multiple reward functions coincide with the observed behavior, and the actual reward function is not identifiable without prior knowledge or supplementary information. Here, we propose Bayesian OptimizationIRL (BOIRL), an IRL framework that identifies multiple solutions that are consistent with the expert demonstrations by efficiently exploring the reward function space. BOIRL achieves this by utilizing Bayesian Optimization along with our newly proposed kernel that (a) projects the parameters of policy invariant reward functions to a single point in a latent space, and (b) ensures that nearby points in the latent space correspond to reward functions that yield similar likelihoods. This projection allows for the use of standard stationary kernels in the latent space to capture the correlations present across the reward function space. Empirical results on synthetic and realworld environments (modelfree and modelbased) show that BOIRL discovers multiple reward functions while minimizing the number of expensive exact policy optimizations.
 Federated Bayesian Optimization via Thompson Sampling.
Zhongxiang Dai^{}, Kian Hsiang Low & Patrick Jaillet^{}.
In Advances in Neural Information Processing Systems 33: 34th Annual Conference on Neural Information Processing Systems (NeurIPS'20), pages 96879699, Dec 612, 2020.
20.1% acceptance rate
Abstract. Bayesian optimization (BO) is a prominent method for optimizing expensivetocompute blackbox functions. The massive computational capability of edge devices such as mobile phones, coupled with privacy concerns, has led to immense recent interest in federated learning (FL), which focuses on collaborative training of deep neural networks (DNN) via firstorder optimization techniques. However, some common machine learning tasks such as hyperparameter tuning of DNN lack access to gradients and thus require zerothorder optimization (blackbox optimization). This hints at the considerable potential of extending BO to the FL setting (FBO), to allow agents to collaborate in these blackbox optimization tasks. Here, we introduce federated Thompson sampling (FTS), which overcomes a number of key challenges of FBO and FL in a principled way: We (a) use random Fourier features to approximate the Gaussian process surrogate model used in BO which naturally produces the parameters to be exchanged between agents, (b) design FTS based on Thompson sampling which significantly reduces the number of parameters to be exchanged, and (c) provide a theoretical convergence guarantee that is robust against heterogeneous agents which is a major challenge in FL and FBO. We empirically demonstrate the effectiveness of FTS in terms of communication efficiency, computational efficiency and practical performance.
 Variational Bayesian Unlearning.
Quoc Phong Nguyen^{}, Kian Hsiang Low & Patrick Jaillet^{}.
In Advances in Neural Information Processing Systems 33: 34th Annual Conference on Neural Information Processing Systems (NeurIPS'20), pages 1602516036, Dec 612, 2020.
20.1% acceptance rate
Abstract. This paper studies the problem of approximately unlearning a Bayesian model from a small subset of the training data to be erased. We frame this problem as one of minimizing the KullbackLeibler distance between the approximate posterior belief of model parameters after directly unlearning from the erased data vs. the exact posterior belief from retraining with remaining data. Using the variational inference (VI) framework, we show that it is equivalent to minimizing an evidence upper bound which trades off between fully unlearning from erased data vs. not entirely forgetting the posterior belief given the full data (i.e., including the remaining data); the latter prevents catastrophic unlearning that can render the model useless. In model training with VI, only an approximate (instead of exact) posterior belief given the full data can be obtained, which makes unlearning even more challenging. We propose two novel tricks to tackle this challenge. We empirically demonstrate our unlearning methods on Bayesian models such as sparse Gaussian process and logistic regression using synthetic and realworld datasets.
 Collaborative Machine Learning with IncentiveAware Model Rewards.
Rachael Hwee Ling Sim^{}, Yehong Zhang^{}, Mun Choon Chan^{} & Kian Hsiang Low.
In Proceedings of the 37th International Conference on Machine Learning (ICML20), pages 89278936, Jun 1218, 2020.
21.8% acceptance rate
Abstract. Collaborative machine learning (ML) is an appealing paradigm to build highquality ML models by training on the aggregated data from many parties. However, these parties are only willing to share their data when given enough incentives, such as a guaranteed fair reward based on their contributions. This motivates the need for measuring a party's contribution and designing an incentiveaware reward scheme accordingly. This paper proposes to value a party's reward based on Shapley value and information gain on model parameters given its data. Subsequently, we give each party a model as a reward. To formally incentivize the collaboration, we define some desirable properties (e.g., fairness and stability) which are inspired by cooperative game theory but adapted for our model reward that is uniquely freely replicable. Then, we propose a novel model reward scheme to satisfy fairness and trade off between the desirable properties via an adjustable parameter. The value of each party's model reward determined by our scheme is attained by injecting Gaussian noise to the aggregated training data with an optimized noise variance. We empirically demonstrate interesting properties of our scheme and evaluate its performance using synthetic and realworld datasets.
 R2B2: Recursive ReasoningBased Bayesian Optimization for NoRegret Learning in Games.
Zhongxiang Dai^{}, Yizhou Chen^{}, Kian Hsiang Low, Patrick Jaillet^{} & TeckHua Ho^{}.
In Proceedings of the 37th International Conference on Machine Learning (ICML20), pages 22912301, Jun 1218, 2020.
21.8% acceptance rate
Abstract. This paper presents a recursive reasoning formalism of Bayesian optimization (BO) to model the reasoning process in the interactions between boundedly rational, selfinterested agents with unknown, complex, and costlytoevaluate payoff functions in repeated games, which we call Recursive ReasoningBased BO (R2B2).
Our R2B2 algorithm is general in that it does not constrain the relationship among the payoff functions of different agents and can thus be applied to various types of games such as constantsum, generalsum, and commonpayoff games. We prove that by reasoning at level 2 or more and at one level higher than the other agents, our R2B2 agent can achieve faster asymptotic convergence to no regret than that without utilizing recursive reasoning. We also propose a computationally cheaper variant of R2B2 called R2B2Lite at the expense of a weaker convergence guarantee. The performance and generality of our R2B2 algorithm are empirically demonstrated using synthetic games, adversarial machine learning, and multiagent reinforcement learning.
 Private Outsourced Bayesian Optimization.
Dmitrii Kharkovskii^{}, Zhongxiang Dai^{} & Kian Hsiang Low.
In Proceedings of the 37th International Conference on Machine Learning (ICML20), pages 52315242, Jun 1218, 2020.
21.8% acceptance rate
Abstract. This paper presents the privateoutsourcedGaussian processupper confidence bound (POGPUCB) algorithm, which is the first algorithm for privacypreserving Bayesian optimization (BO) in the outsourced setting with a provable performance guarantee. We consider the outsourced setting where the entity holding the dataset and the entity performing BO are represented by different parties, and the dataset cannot be released nonprivately. For example, a hospital holds a dataset of sensitive medical records and outsources the BO task on this dataset to an industrial AI company.
The key idea of our approach is to make the BO performance of our algorithm similar to that of nonprivate GPUCB run using the original dataset, which is achieved by using a random projectionbased transformation that preserves both privacy and the pairwise distances between inputs. Our main theoretical contribution is to show that a regret bound similar to that of the standard GPUCB algorithm can be established for our POGPUCB algorithm. We empirically evaluate the performance of our POGPUCB algorithm with synthetic and realworld datasets.
 Learning TaskAgnostic Embedding of Multiple BlackBox Experts for MultiTask Model Fusion.
Trong Nghia Hoang^{}, Chi Thanh Lam^{}, Kian Hsiang Low & Patrick Jaillet^{}.
In Proceedings of the 37th International Conference on Machine Learning (ICML20), pages 42824292, Jun 1218, 2020.
21.8% acceptance rate
Abstract. Model fusion is an emerging study in collective learning where heterogeneous experts with private data and learning architectures need to combine their blackbox knowledge for better performance. Existing literature achieves this via a local knowledge distillation scheme that transfuses the predictive patterns of each pretrained expert onto a whitebox imitator model, which can be incorporated efficiently into a global model. This scheme however does not extend to multitask scenarios where different experts were trained to solve different tasks and only part of their distilled knowledge is relevant to a new task. To address this multitask challenge, we develop a new fusion paradigm that represents each expert as a distribution over a spectrum of predictive prototypes, which are isolated from taskspecific information encoded within the prototype distribution. The taskagnostic prototypes can then be reintegrated to generate a new model that solves a new task encoded with a different prototype distribution. The fusion and adaptation performance of the proposed framework is demonstrated empirically on several realworld benchmark datasets.
 Nonmyopic Gaussian Process Optimization with MacroActions.
Dmitrii Kharkovskii^{}, Chun Kai Ling^{} & Kian Hsiang Low.
In Proceedings of the 23rd International Conference on Artificial Intelligence and Statistics (AISTATS20), pages 45934604, Aug 2628, 2020.
28.7% acceptance rate
Abstract. This paper presents a multistaged approach to nonmyopic adaptive Gaussian process optimization (GPO) for Bayesian optimization (BO) of unknown, highly complex objective functions that, in contrast to existing nonmyopic adaptive BO algorithms, exploits the notion of macroactions for scaling up to a further lookahead to match up to a larger available budget. To achieve this, we generalize GP upper confidence bound to a new acquisition function defined w.r.t. a nonmyopic adaptive macroaction policy, which is intractable to be optimized exactly due to an uncountable set of candidate outputs. The contribution of our work here is thus to derive a nonmyopic adaptive ϵBayesoptimal macroaction GPO (ϵMacroGPO) policy. To perform nonmyopic adaptive BO in real time, we then propose an asymptotically optimal anytime variant of our ϵMacroGPO policy with a performance guarantee. We empirically evaluate the performance of our ϵMacroGPO policy and its anytime variant in BO with synthetic and realworld datasets.
 Scalable Variational Bayesian Kernel Selection for Sparse Gaussian Process Regression.
Tong Teng^{}, Jie Chen^{}, Yehong Zhang^{} & Kian Hsiang Low.
In Proceedings of the 34th AAAI Conference on Artificial Intelligence (AAAI20), pages 59976004, New York, NY, Feb 712, 2020.
20.6% acceptance rate
Abstract. This paper presents a variational Bayesian kernel selection (VBKS) algorithm for sparse Gaussian process regression (SGPR) models. In contrast to existing GP kernel selection algorithms that aim to select only one kernel with the highest model evidence, our proposed VBKS algorithm considers the kernel as a random variable and learns its belief from data such that the uncertainty of the kernel can be interpreted and exploited to avoid overconfident GP predictions. To achieve this, we represent the probabilistic kernel as an additional variational variable in a variational inference (VI) framework for SGPR models where its posterior belief is learned together with that of the other variational variables (i.e., inducing variables and kernel hyperparameters). In particular, we transform the discrete kernel belief into a continuous parametric distribution via reparameterization in order to apply VI. Though it is computationally challenging to jointly optimize a large number of hyperparameters due to many kernels being evaluated simultaneously by our VBKS algorithm, we show that the variational lower bound of the logmarginal likelihood can be decomposed into an additive form such that each additive term depends only on a disjoint subset of the variational variables and can thus be optimized independently. Stochastic optimization is then used to maximize the variational lower bound by iteratively improving the variational approximation of the exact posterior belief via stochastic gradient ascent, which incurs constant time per iteration and hence scales to big data. We empirically evaluate the performance of our VBKS algorithm on synthetic and massive realworld datasets.
 Gaussian Process Decentralized Data Fusion Meets Transfer Learning in LargeScale Distributed Cooperative Perception.
Ruofei Ouyang^{} & Kian Hsiang Low.
Autonomous Robots (Special Issue on MultiRobot and MultiAgent Systems), volume 44, issue 3, pages 359376, Mar 2020.
Extended version of our
AAAI18 paper
Abstract. This paper presents novel Gaussian process decentralized data fusion algorithms exploiting the notion of agentcentric support sets for distributed cooperative perception of largescale environmental phenomena. To overcome the limitations of scale in existing works, our proposed algorithms allow every mobile sensing agent to utilize a different support set and dynamically switch to another during execution for encapsulating its own data into a local summary that, perhaps surprisingly, can still be assimilated with the other agents' local summaries (i.e., based on their current choices of support sets) into a globally consistent summary to be used for predicting the phenomenon. To achieve this, we propose a novel transfer learning mechanism for a team of agents capable of sharing and transferring information encapsulated in a summary based on a support set to that utilizing a different support set with some loss that can be theoretically bounded and analyzed. To alleviate the issue of information loss accumulating over multiple instances of transfer learning, we propose a new information sharing mechanism to be incorporated into our algorithms in order to achieve memoryefficient lazy transfer learning. Empirical evaluation on three realworld datasets for up to 128 agents show that our algorithms outperform the stateoftheart methods.
 FCMSketch: Generic Network Measurements with Data Plane Support.
Cha Hwan Song^{}, Pravein Govindan Kannan, Kian Hsiang Low & Mun Choon Chan^{}.
In Proceedings of the 16th International Conference on emerging Networking EXperiments and Technologies (CoNEXT'20), pages 7892, Dec 14, 2020.
24% acceptance rate
Abstract. Sketches have successfully provided accurate and finegrained measurements (e.g., flow size and heavy hitters) which are imperative for network management. In particular, CountMin (CM) sketch is widely utilized in many applications due to its simple design and ease of implementation. There have been many efforts to build monitoring frameworks based on CountMin sketch. However, these frameworks either support very specific measurement tasks or they cannot be implemented on highspeed programmable hardware (PISA). In this work, we propose FCM, a framework that is designed to support generic network measurement with high accuracy. Our key contribution is FCMSketch, a data structure that has a lightweight implementation on the emerging PISA programmable switches. FCMSketch can also be used as a substitute for CMSketch in applications that use CMSketch. We have implemented FCMSketch on a commodity programmable switch (Barefoot Tofino) using the P4 language. Our evaluation shows that FCMSketch can reduce the errors in many measurement tasks by 50% to 80% compared to CMSketch and other stateoftheart approaches.
 Implicit Posterior Variational Inference for Deep Gaussian Processes.
Haibin Yu^{}, Yizhou Chen^{}, Zhongxiang Dai^{}, Kian Hsiang Low & Patrick Jaillet^{}.
In Advances in Neural Information Processing Systems 32: 33rd Annual Conference on Neural Information Processing Systems (NeurIPS'19), pages 1447514486, Vancouver, Canada, Dec 712, 2019.
3% acceptance rate (spotlight presentation)
Abstract. A multilayer deep Gaussian process (DGP) model is a hierarchical composition of GP models with a greater expressive power. Exact DGP inference is intractable, which has motivated the recent development of deterministic and stochastic approximation methods. Unfortunately, the deterministic approximation methods yield a biased posterior belief while the stochastic one is computationally costly. This paper presents an implicit posterior variational inference (IPVI) framework for DGPs that can ideally recover an unbiased posterior belief and still preserve time efficiency. Inspired by generative adversarial networks, our IPVI framework achieves this by casting the DGP inference problem as a twoplayer game in which a Nash equilibrium, interestingly, coincides with an unbiased posterior belief. This consequently inspires us to devise a bestresponse dynamics algorithm to search for a Nash equilibrium (i.e., an unbiased posterior belief). Empirical evaluation shows that IPVI outperforms the stateoftheart approximation methods for DGPs.
 Bayesian Optimization with Binary Auxiliary Information.
Yehong Zhang^{}, Zhongxiang Dai^{} & Kian Hsiang Low.
In Proceedings of the 35th Conference on Uncertainty in Artificial Intelligence (UAI19), pages 12221232, Tel Aviv, Israel, Jul 2225, 2019.
26.2% acceptance rate (plenary talk)
Subsumes our work on InformationBased MultiFidelity Bayesian Optimization presented in
NeurIPS'17 Workshop on Bayesian Optimization, Long Beach, CA, Dec 9, 2017.
Abstract. This paper presents novel mixedtype Bayesian optimization (BO) algorithms to accelerate the optimization of a target objective function by exploiting correlated auxiliary information of binary type that can be more cheaply obtained, such as in policy search for reinforcement learning and hyperparameter tuning of machine learning models with early stopping. To achieve this, we first propose a mixedtype multioutput Gaussian process (MOGP) to jointly model the continuous target function and binary auxiliary functions. Then, we propose informationbased acquisition functions such as mixedtype entropy search (MTES) and mixedtype predictive ES (MTPES) for mixedtype BO based on the MOGP predictive belief of the target and auxiliary functions. The exact acquisition functions of MTES and MTPES cannot be computed in closed form and need to be approximated. We derive an efficient approximation of MTPES via a novel mixedtype random features approximation of the MOGP model whose crosscorrelation structure between the target and auxiliary functions can be exploited for improving the belief of the global target maximizer using the observations from evaluating these functions. We also propose new practical constraints to relate the global target maximizer to the binary auxiliary functions. We empirically evaluate the performance of MTES and MTPES with synthetic and realworld experiments.
 Bayesian Optimization Meets Bayesian Optimal Stopping.
Zhongxiang Dai^{}, Haibin Yu^{}, Kian Hsiang Low & Patrick Jaillet^{}.
In Proceedings of the 36th International Conference on Machine Learning (ICML19), pages 14961506, Long Beach, CA, Jun 915, 2019.
22.6% acceptance rate
Abstract. Bayesian optimization (BO) is a popular paradigm for optimizing the hyperparameters of machine learning (ML) models due to its sample efficiency. Many ML models require running an iterative training procedure (e.g., stochastic gradient descent). This motivates the question whether information available during the training process (e.g., validation accuracy after each epoch) can be exploited for improving the epoch efficiency of BO algorithms by earlystopping model training under hyperparameter settings that will end up underperforming and hence eliminating unnecessary training epochs. This paper proposes to unify BO (specifically, Gaussian processupper confidence bound (GPUCB)) with Bayesian optimal stopping (BOBOS) to boost the epoch efficiency of BO. To achieve this, while GPUCB is sampleefficient in the number of function evaluations, BOS complements it with epoch efficiency for each function evaluation by providing a principled optimal stopping mechanism for early stopping. BOBOS preserves the (asymptotic) noregret performance of GPUCB using our specified choice of BOS parameters that is amenable to an elegant interpretation in terms of the explorationexploitation tradeoff. We empirically evaluate the performance of BOBOS and demonstrate its generality in hyperparameter optimization of ML models and two other interesting applications.
 Collective Model Fusion for Multiple BlackBox Experts.
Quang Minh Hoang^{}, Trong Nghia Hoang^{}, Kian Hsiang Low & Carleton Kingsford.
In Proceedings of the 36th International Conference on Machine Learning (ICML19), pages 27422750, Long Beach, CA, Jun 915, 2019.
22.6% acceptance rate
Abstract. Model fusion is a fundamental problem in collective machine learning (ML) where independent experts with heterogeneous learning architectures are required to combine expertise to improve predictive performance. This is particularly challenging in informationsensitive domains (e.g., medical records in healthcare analytics) where experts do not have access to each other's internal architecture and local data. To address this challenge, this paper presents the first collective model fusion framework for multiple experts with heterogeneous blackbox architectures. The proposed method will enable this by addressing the following key issues of how blackbox experts interact to understand the predictive behaviors of one another; how these understandings can be represented and shared efficiently among themselves; and how the shared understandings can be combined to generate highquality consensus prediction. The performance of the resulting framework is analyzed theoretically and demonstrated empirically on several datasets.
 Collective Online Learning of Gaussian Processes in Massive MultiAgent Systems.
Trong Nghia Hoang^{}, Quang Minh Hoang^{}, Kian Hsiang Low & Jonathan P. How.
In Proceedings of the 33rd AAAI Conference on Artificial Intelligence (AAAI19), pages 78507857, Honolulu, HI, Jan 27Feb 1, 2019.
16.2% acceptance rate (oral presentation)
Abstract. This paper presents a novel Collective Online Learning of Gaussian Processes (COOLGP) framework for enabling a massive number of GP inference agents to simultaneously perform (a) efficient online updates of their GP models using their local streaming data with varying correlation structures and (b) decentralized fusion of their resulting online GP models with different learned hyperparameter settings and inducing inputs. To realize this, we exploit the notion of a common encoding structure to encapsulate the local streaming data gathered by any GP inference agent into summary statistics based on our proposed representation, which is amenable to both an efficient online update via an importance sampling trick as well as multiagent model fusion via decentralized message passing that can exploit sparse connectivity among agents for improving efficiency and enhance the robustness of our framework against transmission loss. We provide a rigorous theoretical analysis of the approximation loss arising from our proposed representation to achieve efficient online updates and model fusion. Empirical evaluations show that COOLGP is highly effective in model fusion, resilient to information disparity between agents, robust to transmission loss, and can scale to thousands of agents.
 Towards Robust ResNet: A Small Step but a Giant Leap.
Jingfeng Zhang^{}, Bo Han,
Laura Wynter, Kian Hsiang Low & Mohan Kankanhalli^{}.
In Proceedings of the 28th International Joint Conference on Artificial Intelligence (IJCAI19), pages 42854291, Macao, Aug 1016, 2019.
17.9% acceptance rate
Abstract. This paper presents a simple yet principled approach to boosting the robustness of the residual network (ResNet) that is motivated by the dynamical system perspective. Namely, a deep neural network can be interpreted using a partial differential equation, which naturally inspires us to characterize ResNet by an explicit Euler method. Our analytical studies reveal that the step factor h in the Euler method is able to control the robustness of ResNet in both its training and generalization. Specifically, we prove that a small step factor h can benefit the training robustness for backpropagation; from the view of forwardpropagation, a small h can aid in the robustness of the model generalization. A comprehensive empirical evaluation on both vision CIFAR10 and text AGNEWS datasets confirms that a small h aids both the training and generalization robustness.
 GEE: A Gradientbased Explainable Variational Autoencoder for Network Anomaly Detection.
Quoc Phong Nguyen^{}, Kar Wai Lim^{}, Dinil Mon Divakaran, Kian Hsiang Low & Mun Choon Chan^{}.
In Proceedings of the IEEE Conference on Communications and Network Security (CNS'19), pages 9199, Washington, DC, Jun 1012, 2019.
27.8% acceptance rate
Abstract. This paper looks into the problem of detecting network anomalies by analyzing NetFlow records. While many previous works have used statistical models and machine learning techniques in a supervised way, such solutions have the limitations that they require large amount of labeled data for training and are unlikely to detect zeroday attacks. Existing anomaly detection solutions also do not provide an easy way to explain or identify attacks in the anomalous traffic. To address these limitations, we develop and present GEE, a framework for detecting and explaining anomalies in network traffic. GEE comprises of two components: (i) Variational Autoencoder (VAE) — an unsupervised deeplearning technique for detecting anomalies, and (ii) a gradientbased fingerprinting technique for explaining anomalies. Evaluation of GEE on UGR dataset demonstrates that our approach is effective in detecting different anomalies as well as identifying fingerprints that are good representations of these various attacks.
 Stochastic Variational Inference for Bayesian Sparse Gaussian Process Regression.
Haibin Yu^{}, Trong Nghia Hoang^{}, Kian Hsiang Low & Patrick Jaillet^{}.
In Proceedings of the International Joint Conference on Neural Networks (IJCNN'19), Budapest, Hungary, Jul 1419, 2019.
52.4% acceptance rate
Abstract. This paper presents a novel variational inference framework for deriving a family of Bayesian sparse Gaussian process regression (SGPR) models whose approximations are variationally optimal with respect to the fullrank GPR model enriched with various corresponding correlation structures of the observation noises. Our variational Bayesian SGPR (VBSGPR) models jointly treat both the distributions of the inducing variables and hyperparameters as variational parameters, which enables the decomposability of the variational lower bound that in turn can be exploited for stochastic optimization. Such a stochastic optimization involves iteratively following the stochastic gradient of the variational lower bound to improve its estimates of the optimal variational distributions of the inducing variables and hyperparameters (and hence the predictive distribution) of our VBSGPR models and is guaranteed to achieve asymptotic convergence to them. We show that the stochastic gradient is an unbiased estimator of the exact gradient and can be computed in constant time per iteration, hence achieving scalability to big data. We empirically evaluate the performance of our proposed framework on two realworld, massive datasets.
 Decentralized HighDimensional Bayesian Optimization with Factor Graphs.
Trong Nghia Hoang^{}, Quang Minh Hoang^{}, Ruofei Ouyang^{} & Kian Hsiang Low.
In Proceedings of the 32nd AAAI Conference on Artificial Intelligence (AAAI18), pages 32313238, New Orleans, LA, Feb 28, 2018.
24.55% acceptance rate
Abstract. This paper presents a novel decentralized highdimensional Bayesian optimization (DECHBO) algorithm that, in contrast to existing HBO algorithms, can exploit the interdependent effects of various input components on the output of the unknown objective function f for boosting the BO performance and still preserve scalability in the number of input dimensions without requiring prior knowledge or the existence of a low (effective) dimension of the input space. To realize this, we propose a sparse yet rich factor graph representation of f to be exploited for designing an acquisition function that can be similarly represented by a sparse factor graph and hence be efficiently optimized in a decentralized manner using distributed message passing. Despite richly characterizing the interdependent effects of the input components on the output of f with a factor graph, DECHBO can still guarantee (asymptotic) noregret performance. Empirical evaluation on synthetic and realworld experiments shows that DECHBO outperforms the stateoftheart HBO algorithms.
 Gaussian Process Decentralized Data Fusion Meets Transfer Learning in LargeScale Distributed Cooperative Perception.
Ruofei Ouyang^{} & Kian Hsiang Low.
In Proceedings of the 32nd AAAI Conference on Artificial Intelligence (AAAI18), pages 38763883, New Orleans, LA, Feb 28, 2018.
24.55% acceptance rate
Abstract. This paper presents novel Gaussian process decentralized data fusion algorithms exploiting the notion of agentcentric support sets for distributed cooperative perception of largescale environmental phenomena. To overcome the limitations of scale in existing works, our proposed algorithms allow every mobile sensing agent to choose a different support set and dynamically switch to another during execution for encapsulating its own data into a local summary that, perhaps surprisingly, can still be assimilated with the other agents' local summaries (i.e., based on their current choices of support sets) into a globally consistent summary to be used for predicting the phenomenon. To achieve this, we propose a novel transfer learning mechanism for a team of agents capable of sharing and transferring information encapsulated in a summary based on a support set to that utilizing a different support set with some loss that can be theoretically bounded and analyzed. To alleviate the issue of information loss accumulating over multiple instances of transfer learning, we propose a new information sharing mechanism to be incorporated into our algorithms in order to achieve memoryefficient lazy transfer learning. Empirical evaluation on realworld datasets show that our algorithms outperform the stateoftheart methods.
 Artificial Intelligence Research in Singapore: Assisting the Development of a Smart Nation.
Pradeep Varakantham, Bo An, Bryan Low & Jie Zhang.
AI Magazine, volume 38, issue 3, pages 102105, Fall 2017.
Abstract. Artificial Intelligence (AI) research in Singapore is focused on accelerating the country’s development into a Smart Nation. Specifically, AI has been employed extensively in either augmenting the intelligence of humans or in developing automated methods and systems to improve quality of life in Singapore.
 Distributed Batch Gaussian Process Optimization.
Erik Daxberger^{} & Kian Hsiang Low.
In Proceedings of the 34th International Conference on Machine Learning (ICML17), pages 951960, Sydney, Australia, Aug 611, 2017.
25.9% acceptance rate
Abstract. This paper presents a novel distributed batch Gaussian process upper confidence bound (DBGPUCB) algorithm for performing batch Bayesian optimization (BO) of highly complex, costlytoevaluate blackbox objective functions. In contrast to existing batch BO algorithms, DBGPUCB can jointly optimize a batch of inputs (as opposed to selecting the inputs of a batch one at a time) while still preserving scalability in the batch size. To realize this, we generalize GPUCB to a new batch variant amenable to a Markov approximation, which can then be naturally formulated as a multiagent distributed constraint optimization problem in order to fully exploit the efficiency of its stateoftheart solvers for achieving linear time in the batch size. Our DBGPUCB algorithm offers practitioners the flexibility to trade off between the approximation quality and time efficiency by varying the Markov order. We provide a theoretical guarantee for the convergence rate of DBGPUCB via bounds on its cumulative regret. Empirical evaluation on synthetic benchmark objective functions and a realworld optimization problem shows that DBGPUCB outperforms the stateoftheart batch BO algorithms.
 A Generalized Stochastic Variational Bayesian Hyperparameter Learning Framework for Sparse Spectrum Gaussian Process Regression.
Quang Minh Hoang^{}, Trong Nghia Hoang^{} & Kian Hsiang Low.
In Proceedings of the 31st AAAI Conference on Artificial Intelligence (AAAI17), pages 20072014, San Francisco, CA, Feb 49, 2017.
24.6% acceptance rate (oral presentation)
Abstract. While much research effort has been dedicated to scaling up sparse Gaussian process (GP) models based on inducing variables for big data, little attention is afforded to the other less explored class of lowrank GP approximations that exploit the sparse spectral representation of a GP kernel. This paper presents such an effort to advance the state of the art of sparse spectrum GP models to achieve competitive predictive performance for massive datasets. Our generalized framework of stochastic variational Bayesian sparse spectrum GP (sVBSSGP) models addresses their shortcomings by adopting a Bayesian treatment of the spectral frequencies to avoid overfitting, modeling these frequencies jointly in its variational distribution to enable their interaction a posteriori, and exploiting local data for boosting the predictive performance. However, such structural improvements result in a variational lower bound that is intractable to be optimized. To resolve this, we exploit a variational parameterization trick to make it amenable to stochastic optimization. Interestingly, the resulting stochastic gradient has a linearly decomposable structure that can be exploited to refine our stochastic optimization method to incur constant time per iteration while preserving its property of being an unbiased estimator of the exact gradient of the variational lower bound. Empirical evaluation on realworld datasets shows that sVBSSGP outperforms stateoftheart stochastic implementations of sparse GP models.
 A Distributed Variational Inference Framework for Unifying Parallel Sparse Gaussian Process Regression Models.
Trong Nghia Hoang^{}, Quang Minh Hoang^{} & Kian Hsiang Low.
In Proceedings of the 33rd International Conference on Machine Learning (ICML16), pages 382391, New York City, NY, Jun 1924, 2016.
24.3% acceptance rate
Abstract. This paper presents a novel distributed variational inference framework that unifies many parallel sparse Gaussian process regression (SGPR) models for scalable hyperparameter learning with big data. To achieve this, our framework exploits a structure of correlated noise process model that represents the observation noises as a finite realization of a highorder Gaussian Markov random process. By varying the Markov order and covariance function for the noise process model, different variational SGPR models result. This consequently allows the correlation structure of the noise process model to be characterized for which a particular variational SGPR model is optimal. We empirically evaluate the predictive performance and scalability of the distributed variational SGPR models unified by our framework on two realworld datasets.
 NearOptimal Active Learning of MultiOutput Gaussian Processes.
Yehong Zhang^{}, Trong Nghia Hoang^{}, Kian Hsiang Low & Mohan Kankanhalli^{}.
In Proceedings of the 30th AAAI Conference on Artificial Intelligence (AAAI16), pages 23512357, Phoenix, AZ, Feb 1217, 2016.
25.75% acceptance rate
Abstract. This paper addresses the problem of active learning of a multioutput Gaussian process (MOGP) model representing multiple types of coexisting correlated environmental phenomena. In contrast to existing works, our active learning problem involves selecting not just the most informative sampling locations to be observed but also the types of measurements at each selected location for minimizing the predictive uncertainty (i.e., posterior joint entropy) of a target phenomenon of interest given a sampling budget. Unfortunately, such an entropy criterion scales poorly in the numbers of candidate sampling locations and selected observations when optimized. To resolve this issue, we first exploit a structure common to sparse MOGP models for deriving a novel active learning criterion. Then, we exploit a relaxed form of submodularity property of our new criterion for devising a polynomialtime approximation algorithm that guarantees a constantfactor approximation of that achieved by the optimal set of selected observations. Empirical evaluation on realworld datasets shows that our proposed approach outperforms existing algorithms for active learning of MOGP and singleoutput GP models.
 Gaussian Process Planning with Lipschitz Continuous Reward Functions: Towards Unifying Bayesian Optimization, Active Learning, and Beyond.
Chun Kai Ling^{}, Kian Hsiang Low & Patrick Jaillet^{}.
In Proceedings of the 30th AAAI Conference on Artificial Intelligence (AAAI16), pages 18601866, Phoenix, AZ, Feb 1217, 2016.
25.75% acceptance rate
Abstract. This paper presents a novel nonmyopic adaptive Gaussian process planning (GPP) framework endowed with a general class of Lipschitz continuous reward functions that can unify some active learning/sensing and Bayesian optimization criteria and offer practitioners some flexibility to specify their desired choices for defining new tasks/problems. In particular, it utilizes a principled Bayesian sequential decision problem framework for jointly and naturally optimizing the explorationexploitation tradeoff. In general, the resulting induced GPP policy cannot be derived exactly due to an uncountable set of candidate observations. A key contribution of our work here thus lies in exploiting the Lipschitz continuity of the reward functions to solve for a nonmyopic adaptive ϵoptimal GPP (ϵGPP) policy. To plan in real time, we further propose an asymptotically optimal, branchandbound anytime variant of ϵGPP with performance guarantee. We empirically demonstrate the effectiveness of our ϵGPP policy and its anytime variant in Bayesian optimization and an energy harvesting task.
 DrMAD: Distilling ReverseMode Automatic Differentiation for Optimizing Hyperparameters of Deep Neural Networks.
Jie Fu, Hongyin Luo, Jiashi Feng, Kian Hsiang Low & TatSeng Chua.
In Proceedings of the 25th International Joint Conference on Artificial Intelligence (IJCAI16), pages 14691475, New York City, NY, Jul 915, 2016.
<25% acceptance rate
Abstract.
The performance of deep neural networks is wellknown to be sensitive to the setting of their hyperparameters. Recent advances in reversemode automatic differentiation allow for optimizing hyperparameters with gradients. The standard way of computing these gradients involves a forward and backward pass of computations. However, the backward pass usually needs to consume unaffordable memory to store all the intermediate variables to exactly reverse the forward training procedure. In this work, we propose a simple but effective method, DrMAD, to distill the knowledge of the forward pass into a shortcut path, through which we approximately reverse the training trajectory. Experiments on two image benchmark datasets show that DrMAD is at least 45 times faster and consumes 100 times less memory compared to stateoftheart methods for optimizing hyperparameters with minimal compromise to its effectiveness. To the best of our knowledge, DrMAD is the first research attempt to make it practical to automatically tune thousands of hyperparameters of deep neural networks.
 MultiAgent Continuous Transportation with Online Balanced Partitioning.
Chao Wang^{}, Somchaya Liemhetcharat^{} & Kian Hsiang Low.
In Proceedings of the
15th International Conference on Autonomous Agents and MultiAgent Systems (AAMAS16), pages 13031304, Singapore, May 913, 2016.
Abstract. We introduce the concept of continuous transportation task to the context of multiagent systems. A continuous transportation task is one in which a multiagent team visits a number of fixed locations, picks up objects, and delivers them to a transportation hub. The goal is to maximize the rate of transportation while the objects are replenished over time. In this extended abstract, we present a hybrid of centralized and distributed approaches that minimize communications in the multiagent team. We contribute a novel online partitioningtransportation algorithm with information gathering in the multiagent team.
 Conceptbased Hybrid Fusion of Multimodal Event Signals.
Yuhui Wang, Christian von der Weth, Yehong Zhang^{}, Kian Hsiang Low, Vivek Singh & Mohan Kankanhalli^{}.
In Proceedings of the
IEEE International Symposium on Multimedia (ISM'16), pages 1419, San Jose, CA, Dec 1113, 2016.
26.1% acceptance rate
Abstract. Recent years have seen a significant increase in the number of sensors and resulting event related sensor data, allowing for a better monitoring and understanding of realworld events and situations. Eventrelated data come from not only physical sensors (e.g., CCTV cameras, webcams) but also social or microblogging platforms (e.g., Twitter). Given the widespread availability of sensors, we observe that sensors of different modalities often independently observe the same events. We argue that fusing multimodal data about an event can be helpful for more accurate detection, localization and detailed description of events of interest. However, multimodal data often include noisy observations, varying information densities and heterogeneous representations, which makes the fusion a challenging task. In this paper, we propose a hybrid fusion approach that takes the spatial and semantic characteristics of sensor signals about events into account. For this, we first adopt the concept of an imagebased representation that expresses the situation of particular visual concepts (e.g., "crowdedness", "people marching") called Cmage for both physical and social sensor data. Based on this Cmage representation, we model sparse sensor information using a Gaussian process, fuses multimodal event signals with a Bayesian approach, and incorporates spatial relations between the sensor and social observations. We demonstrate the effectiveness of our approach as a proofofconcept over realworld data. Our early results show that the proposed approach can reliably reduce the sensorrelated noise, localize event place, improve event detection reliability, and add semantic context so that the fused data provide a better picture of the observed events or situations.
 Inverse Reinforcement Learning with Locally Consistent Reward Functions.
Quoc Phong Nguyen^{}, Kian Hsiang Low & Patrick Jaillet^{}.
In C. Cortes, N. D. Lawrence, D. D. Lee, M. Sugiyama, R. Garnett, editors, Advances in Neural Information Processing Systems 28: 29th Annual Conference on Neural Information Processing Systems (NeurIPS'15), pages 17471755, Curran Associates, Inc., Montreal, Canada, Dec 712, 2015.
21.9% acceptance rate
Abstract. Existing inverse reinforcement learning (IRL) algorithms have assumed each expert's demonstrated trajectory to be produced by only a single reward function. This paper presents a novel generalization of the IRL problem that allows each trajectory to be generated by multiple locally consistent reward functions, hence catering to more realistic and complex experts' behaviors.
Solving our generalized IRL problem thus involves not only learning these reward functions but also the stochastic transitions between them at any state (including unvisited states).
By representing our IRL problem with a probabilistic graphical model, an expectationmaximization (EM) algorithm can be devised to iteratively learn the reward functions and stochastic transitions between them in order to jointly improve the likelihood of the expert's demonstrated trajectories.
As a result, the most likely partition of a trajectory into segments that are generated from different locally consistent reward functions selected by EM can be derived.
Empirical evaluation on synthetic and realworld datasets shows that our IRL algorithm outperforms the stateoftheart EM clustering with maximum likelihood IRL, which is, interestingly, a reduced variant of our approach.
 Gaussian Process Decentralized Data Fusion and Active Sensing for Spatiotemporal Traffic Modeling and Prediction in MobilityonDemand Systems.
Jie Chen^{}, Kian Hsiang Low, Patrick Jaillet^{} & Yujian Yao^{}.
IEEE Transactions on Automation Science and Engineering
(Special Issue on Networked Cooperative Autonomous Systems), volume 12, issue 3, pages 901921, Jul 2015.
Extended version of our UAI12 and
RSS13 papers
Abstract. Mobilityondemand (MoD) systems have recently emerged as a
promising paradigm of oneway vehicle sharing for sustainable personal
urban mobility in densely populated cities. We assume the capability of
a MoD system to be enhanced by deploying robotic shared vehicles that
can autonomously cruise the streets to be hailed by users. A key
challenge of the MoD system is that of realtime, finegrained mobility
demand and traffic flow sensing and prediction. This paper presents
novel Gaussian process (GP) decentralized data fusion and active
sensing algorithms for realtime, finegrained traffic modeling and
prediction with a fleet of MoD vehicles. The predictive performance of
our decentralized data fusion algorithms are theoretically guaranteed to
be equivalent to that of sophisticated centralized sparse GP
approximations. We derive consensus filtering variants requiring only
local communication between neighboring vehicles. We theoretically
guarantee the performance of our decentralized active sensing
algorithms. When they are used to gather informative data for mobility
demand prediction, they can achieve a dual effect of fleet rebalancing
to service mobility demands. Empirical evaluation on realworld datasets
shows that our algorithms are significantly more timeefficient and
scalable in the size of data and fleet while achieving predictive
performance comparable to that of stateoftheart algorithms.
 A Unifying Framework of Anytime Sparse Gaussian Process Regression Models with Stochastic Variational Inference for Big Data.
Trong Nghia Hoang^{}, Quang Minh Hoang^{} & Kian Hsiang Low.
In Proceedings of the 32nd International Conference on Machine Learning (ICML15), pages 569578, Lille, France, Jul 611, 2015.
26.0% acceptance rate
Abstract. This paper presents a novel unifying framework of anytime sparse Gaussian process regression (SGPR) models that can produce good predictive performance fast and improve their predictive performance over time. Our proposed unifying framework reverses the variational inference procedure to theoretically construct a nontrivial, concave functional that is maximized at the predictive distribution of any SGPR model of our choice.
As a result, a stochastic natural gradient ascent method can be derived that involves iteratively following the stochastic natural gradient of the functional to improve its estimate of the predictive distribution of the chosen SGPR model
and is guaranteed to achieve asymptotic convergence to it. Interestingly, we show that if the predictive distribution of the chosen SGPR model
satisfies certain decomposability conditions, then the stochastic natural gradient is an unbiased estimator of the exact natural gradient and can be computed in constant time (i.e., independent of data size) at each iteration. We empirically evaluate the tradeoff between the predictive performance vs. time efficiency of the anytime SGPR models on two realworld millionsized datasets.
 Parallel Gaussian Process Regression for Big Data: LowRank Representation Meets Markov Approximation.
Kian Hsiang Low, Jiangbo Yu^{}, Jie Chen^{} & Patrick Jaillet^{}.
In Proceedings of the 29th AAAI Conference on Artificial Intelligence (AAAI15), pages 28212827, Austin, TX, Jan 2529, 2015.
26.67% acceptance rate
Abstract. The expressive power of a Gaussian process (GP) model comes at a cost of poor scalability in the data size.
To improve its scalability, this paper presents a lowrankcumMarkov approximation (LMA) of the GP model that is novel in leveraging the dual computational advantages stemming from complementing a lowrank approximate representation of the fullrank GP based on a support set of inputs with a Markov approximation of the resulting residual process; the latter approximation is guaranteed to be closest in the KullbackLeibler distance criterion subject to some constraint
and is considerably more refined than that of existing sparse GP models utilizing lowrank representations due to its more relaxed conditional independence assumption (especially with larger data).
As a result, our LMA method can trade off between the size of the support set and the order of the Markov property to (a) incur lower computational cost than such sparse GP models while achieving predictive performance comparable to them and (b) accurately represent features/patterns of any scale.
Interestingly, varying the Markov order produces a spectrum of LMAs
with PIC approximation and fullrank GP at the two extremes.
An advantage of our LMA method is that it is amenable to parallelization on multiple machines/cores, thereby gaining greater scalability.
Empirical evaluation on three realworld datasets in clusters of up to 32 computing nodes shows that our centralized and parallel LMA methods are significantly more timeefficient and scalable than stateoftheart sparse and fullrank GP regression methods
while achieving comparable predictive performances.
 Recent Advances in Scaling up Gaussian Process Predictive Models for Large Spatiotemporal Data.
Kian Hsiang Low, Jie Chen^{}, Trong Nghia Hoang^{}, Nuo Xu^{} & Patrick Jaillet^{}.
In S. Ravela, A. Sandu, editors,
Dynamic DataDriven Environmental Systems Science  First International Conference, DyDESS'14, LNCS 8964, pages 167181, Springer International Publishing, MIT, Cambridge, MA, Nov 57, 2014.
Oral presentation
Abstract. The expressive power of Gaussian process (GP) models comes at a cost of poor scalability in the size of the data. To improve their scalability, this paper presents an overview of our recent progress in scaling up GP models for large spatiotemporally correlated data through parallelization on clusters of machines, online learning, and nonmyopic active sensing/learning.
 MultiAgent Ad Hoc Team Partitioning by Observing and Modeling SingleAgent Performance.
Etkin Baris Ozgul^{}, Somchaya Liemhetcharat^{} & Kian Hsiang Low.
In Proceedings of the
AsiaPacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC'14), pages 17, Siem Reap, city of Angkor Wat, Cambodia, Dec 912, 2014.
Abstract. Multiagent research has focused on finding the optimal team for a task. Many approaches assume that the performance of the agents are known a priori. We are interested in ad hoc teams, where the agents' algorithms and performance are initially unknown. We focus on the task of modeling the performance of single agents through observation in training environments, and using the learned models to partition a new environment for a multiagent team. The goal is to minimize the number of agents used, while maintaining a performance threshold of the multiagent team. We contribute a novel model to learn the agent's performance through observations, and a partitioning algorithm that minimizes the team size. We evaluate our algorithms in simulation, and show the efficacy of our learn model and partitioning algorithm.
 Scalable DecisionTheoretic Coordination and Control for Realtime Active MultiCamera Surveillance.
Prabhu Natarajan^{}, Trong Nghia Hoang^{}, Yongkang Wong, Kian Hsiang Low & Mohan Kankanhalli^{}.
In Proceedings of the
8th ACM/IEEE International Conference on Distributed Smart Cameras (ICDSC'14) (Invited Paper to Special Session on Smart Cameras for Smart Environments), pages 115120, Venezia, Italy, Nov 47, 2014.
Abstract. This paper presents an overview of our novel decisiontheoretic multiagent approach for controlling and coordinating multiple active cameras in surveillance. In this approach, a surveillance task is modeled as a stochastic optimization problem, where the active cameras are controlled and coordinated to achieve the desired surveillance goal in presence of uncertainties. We enumerate the practical issues in active camera surveillance and discuss how these issues are addressed in our decisiontheoretic approach. We focus on two novel surveillance tasks: maximize the number of targets observed in active cameras with guaranteed image resolution and to improve the fairness in observation of multiple targets. We discuss the overview of our novel decisiontheoretic frameworks: Markov decision process and partially observable Markov decision process frameworks for coordinating active cameras in uncertain and partially occluded environments.
 Active Learning is Planning: Nonmyopic ϵBayesOptimal Active Learning of Gaussian Processes.
Trong Nghia Hoang^{}, Kian Hsiang Low, Patrick Jaillet^{} and Mohan Kankanhalli^{}.
In T. Calders, F. Esposito, E. Hüllermeier, R. Meo, editors, Machine Learning and Knowledge Discovery in Databases  European Conference, ECML/PKDD14 Nectar (New Scientific and Technical Advances in Research) Track, Part III, LNCS 8726, pages 494498, Springer Berlin Heidelberg, Nancy, France, Sep 1519, 2014.
Abstract. A fundamental issue in active learning of Gaussian processes is that of the explorationexploitation tradeoff. This paper presents a novel nonmyopic ϵBayesoptimal active learning (ϵBAL) approach that jointly optimizes the tradeoff. In contrast, existing works have primarily developed greedy algorithms or performed exploration and exploitation separately. To perform active learning in real time, we then propose an anytime algorithm based on ϵBAL with performance guarantee and empirically demonstrate using a realworld dataset that, with limited budget, it outperforms the stateoftheart algorithms.
 Generalized Online Sparse Gaussian Processes with Application to Persistent Mobile Robot Localization.
Kian Hsiang Low, Nuo Xu^{}, Jie Chen^{}, Keng Kiat Lim^{} & Etkin Baris Ozgul^{}.
In T. Calders, F. Esposito, E. Hüllermeier, R. Meo, editors, Machine Learning and Knowledge Discovery in Databases  European Conference, ECML/PKDD14 Nectar (New Scientific and Technical Advances in Research) Track, Part III, LNCS 8726, pages 499503, Springer Berlin Heidelberg, Nancy, France, Sep 1519, 2014.
Abstract. This paper presents a novel online sparse Gaussian process (GP) approximation method that is capable of achieving constant time and memory (i.e., independent of the size of the data) per time step. We theoretically guarantee its predictive performance to be equivalent to that of a sophisticated offline sparse GP approximation method. We empirically demonstrate the practical feasibility of using our online sparse GP approximation method through a realworld persistent mobile robot localization experiment.
 No One is Left "Unwatched": Fairness in Observation of Crowds of Mobile Targets in Active Camera Surveillance.
Prabhu Natarajan^{}, Kian Hsiang Low & Mohan Kankanhalli^{}.
In Proceedings of the
21st European Conference on Artificial Intelligence (ECAI14), including Prestigious Applications of Intelligent Systems (PAIS14), pages 11551160, Prague, Czech Republic, Aug 1822, 2014.
Abstract. Central to the problem of active multicamera surveillance is the fundamental issue of fairness in the observation of crowds of targets such that no target is "starved" of observation by the cameras for a long time. This paper presents a principled decisiontheoretic multicamera coordination and control (MC^{2}) algorithm called fairMC^{2} that can coordinate and control the active cameras to achieve maxmin fairness in the observation of crowds of targets moving stochastically. Our fairMC^{2} algorithm is novel in demonstrating how (a) the uncertainty in the locations, directions, speeds, and observation times of the targets arising from the stochasticity of their motion can be modeled probabilistically, (b) the notion of fairness in observing targets can be formally realized in the domain of multicamera surveillance for the first time by exploiting the maxmin fairness metric to formalize our surveillance objective, that is, to maximize the expected minimum observation time over all targets while guaranteeing a predefined image resolution of observing them, and (c) a structural assumption in the state transition dynamics of a surveillance environment can be exploited to improve its scalability to linear time in the number of targets to be observed during surveillance. Empirical evaluation through extensive simulations in realistic surveillance environments shows that fairMC^{2} outperforms the stateoftheart and baseline MC^{2} algorithms. We have also demonstrated the feasibility of deploying our fairMC^{2} algorithm on real AXIS 214 PTZ cameras.
 GPLocalize: Persistent Mobile Robot Localization using Online Sparse Gaussian Process Observation Model.
Nuo Xu^{}, Kian Hsiang Low, Jie Chen^{}, Keng Kiat Lim^{} & Etkin Baris Ozgul^{}.
In Proceedings of the 28th AAAI Conference on Artificial Intelligence (AAAI14), pages 25852592, Quebec City, Canada, Jul 2731, 2014.
16.6% acceptance rate (oral presentation)
Also appeared in
RSS14 Workshop on NonParametric Learning in Robotics, Berkeley, CA, Jul 12, 2014.
Abstract. Central to robot exploration and mapping is the task of persistent localization in environmental fields characterized by spatially correlated measurements. This paper presents a Gaussian process localization (GPLocalize) algorithm that, in contrast to existing works, can exploit the spatially correlated field measurements taken during a robot's exploration (instead of relying on prior training data) for efficiently and scalably learning the GP observation model online through our proposed novel online sparse GP. As a result, GPLocalize is capable of achieving constant time and memory (i.e., independent of the size of the data) per filtering step, which demonstrates the practical feasibility of using GPs for persistent robot localization and autonomy. Empirical evaluation via simulated experiments with realworld datasets and a real robot experiment shows that GPLocalize outperforms existing GP localization algorithms.
 Nonmyopic ϵBayesOptimal Active Learning of Gaussian Processes.
Trong Nghia Hoang^{}, Kian Hsiang Low, Patrick Jaillet^{} and Mohan Kankanhalli^{}.
In Proceedings of the 31st International Conference on Machine Learning (ICML14), pages 739747, Beijing, China, Jun 2126, 2014.
22.4% acceptance rate (cycle 2)
Also appeared in
RSS14 Workshop on NonParametric Learning in Robotics, Berkeley, CA, Jul 12, 2014.
Abstract. A fundamental issue in active learning of Gaussian processes is that of the explorationexploitation tradeoff.
This paper presents a novel nonmyopic ϵBayesoptimal active learning (ϵBAL) approach that jointly and naturally optimizes the tradeoff.
In contrast, existing works have primarily developed myopic/greedy algorithms or performed exploration and exploitation separately.
To perform active learning in real time, we then propose an anytime algorithm based on ϵBAL with performance guarantee and empirically demonstrate using synthetic and realworld datasets that, with limited budget, it outperforms the stateoftheart algorithms.
 MultiRobot Active Sensing of NonStationary Gaussian ProcessBased Environmental Phenomena.
Ruofei Ouyang^{}, Kian Hsiang Low, Jie Chen^{} & Patrick Jaillet^{}.
In Proceedings of the
13th International Conference on Autonomous Agents and MultiAgent Systems (AAMAS14), pages 573580, Paris, France, May 59, 2014.
23.8% acceptance rate
Also appeared in
RSS14 Workshop on NonParametric Learning in Robotics, Berkeley, CA, Jul 12, 2014.
Abstract. A key challenge of environmental sensing and monitoring is that of sensing, modeling, and predicting largescale, spatially correlated environmental phenomena, especially when they are unknown and nonstationary.
This paper presents a decentralized multirobot active sensing (DECMAS) algorithm that can efficiently coordinate the exploration of multiple robots to gather the most informative observations for predicting an unknown, nonstationary phenomenon.
By modeling the phenomenon using a Dirichlet process mixture of Gaussian processes (DPMGPs), our work here is novel in demonstrating how DPMGPs and its structural properties can be exploited to (a) formalize an active sensing criterion that trades off between gathering the most informative observations for estimating the unknown, nonstationary spatial correlation structure vs. that for predicting the phenomenon given the current, imprecise estimate of the correlation structure, and (b) support efficient decentralized coordination.
We also provide a theoretical performance guarantee for DECMAS and analyze its time complexity.
We empirically demonstrate using two realworld datasets that DECMAS outperforms stateoftheart MAS algorithms.
 DecisionTheoretic Approach to Maximizing Fairness in MultiTarget Observation in MultiCamera Surveillance.
Prabhu Natarajan^{}, Kian Hsiang Low & Mohan Kankanhalli^{}.
In Proceedings of the
13th International Conference on Autonomous Agents and MultiAgent Systems (AAMAS14), pages 15211522, Paris, France, May 59, 2014.
Abstract. Central to the problem of active multicamera surveillance is the fundamental issue of fairness in the observation of multiple targets such that no target is left unobserved by the cameras for a long time. To address this important issue, we propose a novel principled decisiontheoretic approach to control and coordinate multiple active cameras to achieve fairness in the observation of multiple moving targets.
 Interactive POMDP Lite: Towards Practical Planning to Predict and Exploit Intentions for Interacting with SelfInterested Agents.
Trong Nghia Hoang^{} & Kian Hsiang Low.
In Proceedings of the 23rd International Joint Conference on Artificial Intelligence (IJCAI13), pages 22982305, Beijing, China, Aug 39, 2013.
13.2% acceptance rate (oral presentation)
Abstract. A key challenge in noncooperative multiagent systems is that of developing efficient planning algorithms for intelligent agents to interact and perform effectively among boundedly rational, selfinterested agents (e.g., humans). The practicality of existing works addressing this challenge is being undermined due to either the restrictive assumptions of the other agents' behavior, the failure in accounting for their rationality, or the prohibitively expensive cost of modeling and predicting their intentions. To boost the practicality of research in this field, we investigate how intention prediction can be efficiently exploited and made practical in planning, thereby leading to efficient intentionaware planning frameworks capable of predicting the intentions of other agents and acting optimally with respect to their predicted intentions. We show that the performance losses incurred by the resulting planning policies are linearly bounded by the error of intention prediction. Empirical evaluations through a series of stochastic games demonstrate that our policies can achieve better and more robust performance than the stateoftheart algorithms.
 A General Framework for Interacting BayesOptimally with SelfInterested Agents using Arbitrary Parametric Model and Model Prior.
Trong Nghia Hoang^{} & Kian Hsiang Low.
In Proceedings of the 23rd International Joint Conference on Artificial Intelligence (IJCAI13), pages 13941400, Beijing, China, Aug 39, 2013.
28.0% acceptance rate
Abstract. Recent advances in Bayesian reinforcement learning (BRL) have shown that Bayesoptimality is theoretically achievable by modeling the environment's latent dynamics using FlatDirichletMultinomial (FDM) prior. In selfinterested multiagent environments, the transition dynamics are mainly controlled by the other agent's stochastic behavior for which FDM's independence and modeling assumptions do not hold. As a result, FDM does not allow the other agent's behavior to be generalized across different states nor specified using prior domain knowledge. To overcome these practical limitations of FDM, we propose a generalization of BRL to integrate the general class of parametric models and model priors, thus allowing practitioners' domain knowledge to be exploited to produce a finegrained and compact representation of the other agent's behavior. Empirical evaluation shows that our approach outperforms existing multiagent reinforcement learning algorithms.
 Parallel Gaussian Process Regression with LowRank Covariance Matrix Approximations.
Jie Chen^{}, Nannan Cao^{}, Kian Hsiang Low, Ruofei Ouyang^{}, Colin KengYan Tan & Patrick Jaillet^{}.
In Proceedings of the 29th Conference on Uncertainty in Artificial Intelligence (UAI13), pages 152161, Bellevue, WA, Jul 1115, 2013.
31.3% acceptance rate
Abstract. Gaussian processes (GP) are Bayesian nonparametric models that are widely used for probabilistic regression. Unfortunately, it cannot scale well with large data nor perform realtime predictions due to its cubic time cost in the data size. This paper presents two parallel GP regression methods that exploit lowrank covariance matrix approximations for distributing the computational load among parallel machines to achieve time efficiency and scalability. We theoretically guarantee the predictive performances of our proposed parallel GPs to be equivalent to that of some centralized approximate GP regression methods: The computation of their centralized counterparts can be distributed among parallel machines, hence achieving greater time efficiency and scalability. We analytically compare the properties of our parallel GPs such as time, space, and communication complexity. Empirical evaluation on two realworld datasets in a cluster of 20 computing nodes shows that our parallel GPs are significantly more timeefficient and scalable than their centralized counterparts and exact/full GP while achieving predictive performances comparable to full GP.
 Gaussian ProcessBased Decentralized Data Fusion and Active Sensing for MobilityonDemand System.
Jie Chen^{}, Kian Hsiang Low & Colin KengYan Tan.
In Proceedings of the
Robotics: Science and Systems Conference (RSS13), Berlin, Germany, Jun 2428, 2013.
30.1% acceptance rate
Abstract. Mobilityondemand (MoD) systems have recently emerged as a promising paradigm of oneway vehicle sharing for sustainable personal urban mobility in densely populated cities. In this paper, we enhance the capability of a MoD system by deploying robotic shared vehicles that can autonomously cruise the streets to be hailed by users. A key challenge to managing the MoD system effectively is that of realtime, finegrained mobility demand sensing and prediction. This paper presents a novel decentralized data fusion and active sensing algorithm for realtime, finegrained mobility demand sensing and prediction with a fleet of autonomous robotic vehicles in a MoD system. Our Gaussian process (GP)based decentralized data fusion algorithm can achieve a fine balance between predictive power and time efficiency. We theoretically guarantee its predictive performance to be equivalent to that of a sophisticated centralized sparse approximation for the GP model: The computation of such a sparse approximate GP model can thus be distributed among the MoD vehicles, hence achieving efficient and scalable demand prediction. Though our decentralized active sensing strategy is devised to gather the most informative demand data for demand prediction, it can achieve a dual effect of fleet rebalancing to service the mobility demands. Empirical evaluation on realworld mobility demand data shows that our proposed algorithm can achieve a better balance between predictive accuracy and time efficiency than stateoftheart algorithms.
 MultiRobot Informative Path Planning for Active Sensing of Environmental Phenomena: A Tale of Two Algorithms.
Nannan Cao^{}, Kian Hsiang Low & John M. Dolan^{}.
In Proceedings of the
12th International Conference on Autonomous Agents and MultiAgent Systems (AAMAS13), pages 714, Saint Paul, MN, May 610, 2013.
22.9% acceptance rate
Abstract. A key problem of robotic environmental sensing and monitoring is that of active sensing: How can a team of robots plan the most informative observation paths to minimize the uncertainty in modeling and predicting an environmental phenomenon? This paper presents two principled approaches to efficient informationtheoretic path planning based on entropy and mutual information criteria for in situ active sensing of an important broad class of widelyoccurring environmental phenomena called anisotropic fields. Our proposed algorithms are novel in addressing a tradeoff between active sensing performance and time efficiency. An important practical consequence is that our algorithms can exploit the spatial correlation structure of Gaussian processbased anisotropic fields to improve time efficiency while preserving nearoptimal active sensing performance. We analyze the time complexity of our algorithms and prove analytically that they scale better than stateoftheart algorithms with increasing planning horizon length. We provide theoretical guarantees on the active sensing performance of our algorithms for a class of exploration tasks called transect sampling, which, in particular, can be improved with longer planning time and/or lower spatial correlation along the transect. Empirical evaluation on realworld anisotropic field data shows that our algorithms can perform better or at least as well as the stateoftheart algorithms while often incurring a few orders of magnitude less computational time, even when the field conditions are less favorable.
 Adaptive Sampling of Time Series with Application to Remote Exploration.
David R. Thompson^{}, Nathalie Cabrol, Michael Furlong, Craig Hardgrove, Kian Hsiang Low, Jeffrey Moersch & David Wettergreen.
In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA'13), pages 34633468, Karlsruhe, Germany, May 610, 2013.
Abstract. We address the problem of adaptive informationoptimal data collection in time series. Here a remote sensor or explorer agent throttles its sampling rate in order to track anomalous events while obeying constraints on time and power. This problem is challenging because the agent has limited visibility  all collected datapoints lie in the past, but its resource allocation decisions require predicting far into the future. Our solution is to continually fit a Gaussian process model to the latest data and optimize the sampling plan on line to maximize information gain. We compare the performance characteristics of stationary and nonstationary Gaussian process models. We also describe an application based on geologic analysis during planetary rover exploration. Here adaptive sampling can improve coverage of localized anomalies and potentially benefit mission science yield of long autonomous traverses.
 Decentralized Data Fusion and Active Sensing with Mobile Sensors for Modeling and Predicting Spatiotemporal Traffic Phenomena.
Jie Chen^{}, Kian Hsiang Low, Colin KengYan Tan, Ali Oran^{}, Patrick Jaillet^{}, John M. Dolan^{} & Gaurav S. Sukhatme^{}.
In Proceedings of the 28th Conference on Uncertainty in Artificial Intelligence (UAI12), pages 163173, Catalina Island, CA, Aug 1517, 2012.
31.6% acceptance rate
Also appeared in AAMAS12 Workshop on Agents in Traffic and Transportation (ATT12), Valencia, Spain, June 48, 2012.
Abstract. The problem of modeling and predicting spatiotemporal traffic phenomena over an urban road network is important to many traffic applications such as detecting and forecasting congestion hotspots. This paper presents a decentralized data fusion and active sensing (D2FAS) algorithm for mobile sensors to actively explore the road network to gather and assimilate the most informative data for predicting the traffic phenomenon. We analyze the time and communication complexity of D2FAS and demonstrate that it can scale well with a large number of observations and sensors. We provide a theoretical guarantee on its predictive performance to be equivalent to that of a sophisticated centralized sparse approximation for the Gaussian process (GP) model: The computation of such a sparse approximate GP model can thus be parallelized and distributed among the mobile sensors (in a Googlelike MapReduce paradigm), thereby achieving efficient and scalable prediction. We also theoretically guarantee its active sensing performance that improves under various practical environmental conditions. Empirical evaluation on realworld urban road network data shows that our D2FAS algorithm is significantly more timeefficient and scalable than stateoftheart centralized algorithms while achieving comparable predictive performance.
 Hierarchical Bayesian Nonparametric Approach to Modeling and Learning the Wisdom of Crowds of Urban Traffic Route Planning Agents.
Jiangbo Yu^{}, Kian Hsiang Low, Ali Oran^{} & Patrick Jaillet^{}.
In Proceedings of the IEEE/WIC/ACM International Conference on Intelligent Agent Technology (IAT'12)
(Invited Paper to Special Session on LargeScale ApplicationFocused MultiAgent Systems), pages 478485, Macau, Dec 47, 2012.
Abstract. Route prediction is important to analyzing and understanding the route patterns and behavior of traffic crowds. Its objective is to predict the most likely or "popular" route of road segments from a given point in a road network. This paper presents a hierarchical Bayesian nonparametric approach to efficient and scalable route prediction that can harness the wisdom of crowds of route planning agents by aggregating their sequential routes of possibly varying lengths and origindestination pairs. In particular, our approach has the advantages of (a) not requiring a Markov assumption to be imposed and (b) generalizing well with sparse data, thus resulting in significantly improved prediction accuracy, as demonstrated empirically using realworld taxi route data. We also show two practical applications of our route prediction algorithm: predictive taxi ranking and route recommendation.
 DecisionTheoretic Coordination and Control for Active MultiCamera Surveillance in Uncertain, Partially Observable Environments.
Prabhu Natarajan^{}, Trong Nghia Hoang^{}, Kian Hsiang Low & Mohan Kankanhalli^{}.
In Proceedings of the
6th ACM/IEEE International Conference on Distributed Smart Cameras (ICDSC'12), pages 16, Hong Kong, Oct 30  Nov 2, 2012.
Abstract. A central problem of surveillance is to monitor multiple targets moving in a largescale, obstacleridden environment with occlusions. This paper presents a novel principled Partially Observable Markov Decision Processbased approach to coordinating and controlling a network of active cameras for tracking and observing multiple mobile targets at high resolution in such surveillance environments. Our proposed approach is capable of (a) maintaining a belief over the targets' states (i.e., locations, directions, and velocities) to track them, even when they may not be observed directly by the cameras at all times, (b) coordinating the cameras' actions to simultaneously improve the belief over the targets' states and maximize the expected number of targets observed with a guaranteed resolution, and (c) exploiting the inherent structure of our surveillance problem to improve its scalability (i.e., linear time) in the number of targets to be observed. Quantitative comparisons with stateoftheart multicamera coordination and control techniques show that our approach can achieve higher surveillance quality in real time. The practical feasibility of our approach is also demonstrated using real AXIS 214 PTZ cameras.
 Decentralized Active Robotic Exploration and Mapping for Probabilistic Field Classification in Environmental Sensing.
Kian Hsiang Low, Jie Chen^{}, John M. Dolan^{}, Steve Chien^{} & David R. Thompson^{}.
In Proceedings of the
11th International Conference on Autonomous Agents and MultiAgent Systems (AAMAS12), pages 105112, Valencia, Spain, June 48, 2012.
20.4% acceptance rate
Also appeared in
IROS'11 Workshop on Robotics for Environmental Monitoring (WREM11), San Francisco, CA, Sep 30, 2011.
Abstract. A central problem in environmental sensing and monitoring is to classify/label the hotspots in a largescale environmental field. This paper presents a novel decentralized active robotic exploration (DARE) strategy for probabilistic classification/labeling of hotspots in a Gaussian process (GP)based field. In contrast to existing stateoftheart exploration strategies for learning environmental field maps, the time needed to solve the DARE strategy is independent of the map resolution and the number of robots, thus making it practical for in situ, realtime active sampling. Its exploration behavior exhibits an interesting formal tradeoff between that of boundary tracking until the hotspot region boundary can be accurately predicted and widearea coverage to find new boundaries in sparsely sampled areas to be tracked. We provide a theoretical guarantee on the active exploration performance of the DARE strategy: under reasonable conditional independence assumption, we prove that it can optimally achieve two formal costminimizing exploration objectives based on the misclassification and entropy criteria. Importantly, this result implies that the uncertainty of labeling the hotspots in a GPbased field is greatest at or close to the hotspot region boundaries. Empirical evaluation on realworld plankton density and temperature field data shows that, subject to limited observations, DARE strategy can achieve more superior classification of hotspots and time efficiency than stateoftheart active exploration strategies.
 DecisionTheoretic Approach to Maximizing Observation of Multiple Targets in MultiCamera Surveillance.
Prabhu Natarajan^{}, Trong Nghia Hoang^{}, Kian Hsiang Low & Mohan Kankanhalli^{}.
In Proceedings of the
11th International Conference on Autonomous Agents and MultiAgent Systems (AAMAS12), pages 155162, Valencia, Spain, June 48, 2012.
20.4% acceptance rate
Abstract. This paper presents a novel decisiontheoretic approach to control and coordinate multiple active cameras for observing a number of moving targets in a surveillance system. This approach offers the advantages of being able to (a) account for the stochasticity of targets' motion via probabilistic modeling, and (b) address the tradeoff between maximizing the expected number of observed targets and the resolution of the observed targets through stochastic optimization. One of the key issues faced by existing approaches in multicamera surveillance is that of scalability with increasing number of targets. We show how its scalability can be improved by exploiting the problem structure: as proven analytically, our decisiontheoretic approach incurs time that is linear in the number of targets to be observed during surveillance. As demonstrated empirically through simulations, our proposed approach can achieve highquality surveillance of up to 50 targets in real time and its surveillance performance degrades gracefully with increasing number of targets. We also demonstrate our proposed approach with real AXIS 214 PTZ cameras in maximizing the number of Lego robots observed at high resolution over a surveyed rectangular area. The results are promising and clearly show the feasibility of our decisiontheoretic approach in controlling and coordinating the active cameras in real surveillance system.
 IntentionAware Planning under Uncertainty for Interacting with SelfInterested, Boundedly Rational Agents.
Trong Nghia Hoang^{} & Kian Hsiang Low.
In Proceedings of the
11th International Conference on Autonomous Agents and MultiAgent Systems (AAMAS12), pages 12331234, Valencia, Spain, June 48, 2012.
Abstract. A key challenge in noncooperative multiagent systems is that of developing efficient planning algorithms for intelligent agents to perform effectively among boundedly rational, selfinterested (i.e., noncooperative) agents (e.g., humans). To address this challenge, we investigate how intention prediction can be efficiently exploited and made practical in planning, thereby leading to efficient intentionaware planning frameworks capable of predicting the intentions of other agents and acting optimally with respect to their predicted intentions.
 Active Markov InformationTheoretic Path Planning for Robotic Environmental Sensing.
Kian Hsiang Low, John M. Dolan^{} & Pradeep Khosla^{}.
In Proceedings of the
10th International Conference on Autonomous Agents and MultiAgent Systems (AAMAS11), pages 753760, Taipei, Taiwan, May 26, 2011.
22.1% acceptance rate
Abstract. Recent research in multirobot exploration and mapping has focused on sampling environmental fields, which are typically modeled using the Gaussian process (GP). Existing informationtheoretic exploration strategies for learning GPbased environmental field maps adopt the nonMarkovian problem structure and consequently scale poorly with the length of history of observations. Hence, it becomes computationally impractical to use these strategies for in situ, realtime active sampling. To ease this computational burden, this paper presents a Markovbased approach to efficient informationtheoretic path planning for active sampling of GPbased fields. We analyze the time complexity of solving the Markovbased path planning problem, and demonstrate analytically that it scales better than that of deriving the nonMarkovian strategies with increasing length of planning horizon. For a class of exploration tasks called the transect sampling task, we provide theoretical guarantees on the active sampling performance of our Markovbased policy, from which ideal environmental field conditions and sampling task settings can be established to limit its performance degradation due to violation of the Markov assumption. Empirical evaluation on realworld temperature and plankton density field data shows that our Markovbased policy can generally achieve active sampling performance comparable to that of the widelyused nonMarkovian greedy policies under less favorable realistic field conditions and task settings while enjoying significant computational gain over them.
 Autonomous Personal Vehicle for the First and LastMile Transportation Services.
Zhuang Jie Chong, Baoxing Qin, Tirthankar Bandyopadhyay, Tichakorn Wongpiromsarn, Edward Samuel Rankin, Marcelo H. Ang, Jr.^{}, Emilio Frazzoli^{}, Daniela Rus^{}, David Hsu^{} & Kian Hsiang Low.
In Proceedings of the
5th IEEE International Conference on Cybernetics and
Intelligent Systems and 5th IEEE International
Conference on Robotics, Automation and Mechatronics (CISRAM'11), pages 253260, Qingdao, China, Sep 1719, 2011.
Also appeared in IROS'11 Workshop on Perception and Navigation for Autonomous Vehicles in Human Environment, San Francisco, CA, Sep 30, 2011.
Abstract. This paper describes an autonomous vehicle testbed that aims at providing the first and last mile transportation services. The vehicle mainly operates in a crowded urban environment whose features can be extracted a priori. To ensure that the system is economically feasible, we take a minimalistic approach and exploit prior knowledge of the environment and the availability of the existing infrastructure such as cellular networks and traffic cameras. We present three main components of the system: pedestrian detection, localization (even in the presence of tall buildings) and navigation. The performance of each component is evaluated. Finally, we describe the role of the existing infrastructural sensors and show the improved performance of the system when they are utilized.
 Telesupervised Remote Surface
Water Quality Sensing.
Gregg Podnar, John M. Dolan^{}, Kian Hsiang Low & Alberto Elfes.
In Proceedings of the IEEE Aerospace Conference, Big Sky, MT, Mar 613, 2010.
Abstract. We present a fleet of autonomous Robot Sensor Boats (RSBs) developed for lake and river fresh water quality assessment and controlled by our Multilevel Autonomy Robot Telesupervision Architecture (MARTA). The RSBs are low cost, highly maneuverable, shallow draft sensor boats, developed as part of the Sensor Web program supported under the Advanced Information Systems Technology program of NASA's Earth Systems Technology Office. They can scan large areas of lakes, and navigate up tributaries to measure water quality near outfalls that larger research vessels cannot reach. The MARTA telesupervision architecture has been applied to a number of domains from multiplatform autonomous wide area planetary mineral prospecting, to multiplatform ocean monitoring. The RSBs are a complementary expansion of a fleet of NOAA/NASAdeveloped extendeddeployment surface autonomous vehicles that enable insitu study of meteorological factors of the ocean/atmosphere interface, and which have been adapted to investigate harmful algal blooms under this program. The flexibility of the MARTA telesupervision architecture was proven as it supported simultaneous operation of these heterogenous autonomous sensor platforms while geographically widely separated. Results and analysis are presented of multiple tests carried out over three months using a multisensor water sonde to assess water quality in a small recreational lake. Inference Grids were used to produce maps representing temperature, pH, and dissolved oxygen. The tests were performed under various water conditions (clear vs. hair algaeladen) and both before and after heavy rains. Data from each RSB was relayed to a data server in our lab in Pittsburgh, Pennsylvania, and made available over the World Wide Web where it was acquired by team members at the Jet Propulsion Laboratory of NASA in Pasadena, California who monitored the boats and their sensor readings in real time, as well as using these data to model the water quality by producing Inference Gridbased maps.
 InformationTheoretic Approach to Efficient Adaptive Path Planning for Mobile Robotic Environmental Sensing.
Kian Hsiang Low, John M. Dolan^{} & Pradeep Khosla^{}.
In Proceedings of the 19th International Conference on Automated Planning and Scheduling (ICAPS09), pages 233240, Thessaloniki, Greece, Sep 1923, 2009.
33.9% acceptance rate
Also appeared in IPSN09 Workshop on Sensor Networks for Earth and Space Science Applications (ESSA09), San Francisco, CA, Apr 16, 2009.
Also orally presented in RSS09 Workshop on Aquatic Robots and Ocean Sampling, Seattle, WA, Jun 29, 2009.
Abstract. Recent research in robot exploration and mapping has focused on sampling environmental hotspot fields. This exploration task is formalized by Low, Dolan, and Khosla (2008) in a sequential decisiontheoretic planning under uncertainty framework called MASP. The time complexity of solving MASP approximately depends on the map resolution, which limits its use in largescale, highresolution exploration and mapping. To alleviate this computational difficulty, this paper presents an informationtheoretic approach to MASP (iMASP) for efficient adaptive path planning; by reformulating the costminimizing iMASP as a rewardmaximizing problem, its time complexity becomes independent of map resolution and is less sensitive to increasing robot team size as demonstrated both theoretically and empirically. Using the rewardmaximizing dual, we derive a novel adaptive variant of maximum entropy sampling, thus improving the induced exploration policy performance. It also allows us to establish theoretical bounds quantifying the performance advantage of optimal adaptive over nonadaptive policies and the performance quality of approximately optimal vs. optimal adaptive policies. We show analytically and empirically the superior performance of iMASPbased policies for sampling the logGaussian process to that of policies for the widelyused Gaussian process in mapping the hotspot field. Lastly, we provide sufficient conditions that, when met, guarantee adaptivity has no benefit under an assumed environment model.
 Cooperative Aquatic Sensing using the Telesupervised Adaptive Ocean Sensor Fleet.
John M. Dolan^{}, Gregg W. Podnar, Stephen Stancliff, Kian Hsiang Low, Alberto Elfes, John Higinbotham, Jeffrey C. Hosler, Tiffany A. Moisan & John Moisan.
In Proceedings of the SPIE Conference on Remote Sensing of the Ocean, Sea Ice, and Large Water Regions, volume 7473, Berlin, Germany, Aug 31  Sep 3, 2009.
Abstract. Earth science research must bridge the gap between the atmosphere and the ocean to foster understanding of Earth's climate and ecology. Typical ocean sensing is done with satellites or in situ buoys and research ships which are slow to reposition. Cloud cover inhibits study of localized transient phenomena such as Harmful Algal Blooms (HAB). A fleet of extendeddeployment surface autonomous vehicles will enable in situ study of characteristics of HAB, coastal pollutants, and related phenomena. We have developed a multiplatform telesupervision architecture that supports adaptive reconfiguration based on environmental sensor inputs. Our system allows the autonomous repositioning of smart sensors for HAB study by networking a fleet of NOAA OASIS (Ocean Atmosphere Sensor Integration System) surface autonomous vehicles. In situ measurements intelligently modify the search for areas of high concentration. Inference Grid and complementary informationtheoretic techniques support sensor fusion and analysis. Telesupervision supports sliding autonomy from highlevel mission tasking, through vehicle and data monitoring, to teleoperation when direct human interaction is appropriate. This paper reports on experimental results from multiplatform tests conducted in the Chesapeake Bay and in Pittsburgh, Pennsylvania waters using OASIS platforms, autonomous kayaks, and multiple simulated platforms to conduct cooperative sensing of chlorophylla and water quality.
 Robot Boats as a Mobile Aquatic Sensor Network.
Kian Hsiang Low, Gregg Podnar, Stephen Stancliff, John M. Dolan^{} & Alberto Elfes.
In Proceedings of the IPSN09 Workshop on Sensor Networks for Earth and Space Science Applications (ESSA09), San Francisco, CA, Apr 16, 2009.
Abstract. This paper describes the Multilevel Autonomy Robot Telesupervision Architecture (MARTA), an architecture for supervisory control of a heterogeneous fleet of networked unmanned autonomous aquatic surface vessels carrying a payload of environmental science sensors. This architecture allows a landbased human scientist to effectively supervise data gathering by multiple robotic assets that implement a web of widely dispersed mobile sensors for in situ study of physical, chemical or biological processes in water or in the water/atmosphere interface.
 Adaptive MultiRobot WideArea Exploration And Mapping.
Kian Hsiang Low, John M. Dolan^{} & Pradeep Khosla^{}.
In Proceedings of the
7th International Conference on Autonomous Agents and MultiAgent Systems (AAMAS08), pages 2330, Estoril, Portugal, May 1216, 2008.
22.2% acceptance rate
Also presented as a poster in RSS09 Workshop on Aquatic Robots and Ocean Sampling, Seattle, WA, Jun 29, 2009.
Abstract. The exploration problem is a central issue in mobile robotics. A complete terrain coverage is not practical if the environment is large with only a few small hotspots. This paper presents an adaptive multirobot exploration strategy that is novel in performing both widearea coverage and hotspot sampling using nonmyopic path planning. As a result, the environmental phenomena can be accurately mapped. It is based on a dynamic programming formulation, which we call the Multirobot Adaptive Sampling Problem (MASP). A key feature of MASP is in covering the entire adaptivity spectrum, thus allowing strategies of varying adaptivity to be formed and theoretically analyzed in their performance; a more adaptive strategy improves mapping accuracy. We apply MASP to sampling the Gaussian and logGaussian processes, and analyze if the resulting strategies are adaptive and maximize widearea coverage and hotspot sampling. Solving MASP is nontrivial as it comprises continuous state components. So, it is reformulated for convex analysis, which allows discretestate monotonebounding approximation to be developed. We provide a theoretical guarantee on the policy quality of the approximate MASP (aMASP) for using in MASP. Although aMASP can be solved exactly, its state size grows exponentially with the number of stages. To alleviate this computational difficulty, anytime algorithms are proposed based on aMASP, one of which can guarantee its policy quality for MASP in real time.
 Adaptive Sampling for MultiRobot WideArea Exploration.
Kian Hsiang Low, Geoffrey J. Gordon, John M. Dolan^{} & Pradeep Khosla^{}.
In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA'07), pages 755760, Rome, Italy, Apr 1014, 2007.
Abstract. The exploration problem is a central issue in mobile robotics. A complete coverage is not practical if the environment is large with a few small hotspots, and the sampling cost is high. So, it is desirable to build robot teams that can coordinate to maximize sampling at these hotspots while minimizing resource costs, and consequently learn more accurately about properties of such environmental phenomena. An important issue in designing such teams is the exploration strategy. The contribution of this paper is in the evaluation of an adaptive exploration strategy called adaptive cluster sampling (ACS), which is demonstrated to reduce the resource costs (i.e., mission time and energy consumption) of a robot team, and yield more information about the environment by directing robot exploration towards hotspots. Due to the adaptive nature of the strategy, it is not obvious how the sampled data can be used to provide unbiased, lowvariance estimates of the properties. This paper therefore discusses how estimators that are RaoBlackwellized can be used to achieve low error. This paper also presents the first analysis of the characteristics of the environmental phenomena that favor the ACS strategy and estimators. Quantitative experimental results in a mineral prospecting task simulation show that our approach is more efficient in exploration by yielding more minerals and information with fewer resources and providing more precise mineral density estimates than previous methods.
 Autonomic Mobile Sensor Network with SelfCoordinated Task Allocation and Execution.
Kian Hsiang Low, Wee Kheng Leow^{} & Marcelo H. Ang, Jr.^{}
IEEE Transactions on Systems, Man, and Cybernetics  Part C: Applications and Reviews
(Special Issue on Engineering Autonomic Systems), volume 36, issue 3, pages 315327, May 2006.
Extended version of our IJCAI03, ICRA'04, and
AAAI04 papers
Andrew P. Sage Best Transactions Paper Award for the best paper published in IEEE Trans. SMC  Part A, B, and C in 2006
Abstract. This paper describes a distributed layered architecture for resourceconstrained multirobot cooperation, which is utilized in autonomic mobile sensor network coverage. In the upper layer, a dynamic task allocation scheme selforganizes the robot coalitions to track efficiently across regions. It uses concepts of ant behavior to selfregulate the regional distributions of robots in proportion to that of the moving targets to be tracked in a nonstationary environment. As a result, the adverse effects of task interference between robots are minimized and network coverage is improved. In the lower task execution layer, the robots use selforganizing neural networks to coordinate their target tracking within a region. Both layers employ selforganization techniques, which exhibit autonomic properties such as selfconfiguring, selfoptimizing, selfhealing, and selfprotecting. Quantitative comparisons with other tracking strategies such as static sensor placements, potential fields, and auctionbased negotiation show that our layered approach can provide better coverage, greater robustness to sensor failures, and greater flexibility to respond to environmental changes.
 An Ensemble of Cooperative Extended Kohonen Maps for Complex Robot Motion Tasks.
Kian Hsiang Low, Wee Kheng Leow^{} & Marcelo H. Ang, Jr.^{}
Neural Computation, volume 17, issue 6, pages 14111445, Jun 2005.
Extended version of our AAMAS02, ICRA'03, and
IJCAI03 papers
Abstract. Selforganizing feature maps such as extended Kohonen maps (EKMs) have been very successful at learning sensorimotor control for mobile robot tasks. This letter presents a new ensemble approach, cooperative EKMs with indirect mapping, to achieve complex robot motion. An indirectmapping EKM selforganizes to map from the sensory input space to the motor control space indirectly via a control parameter space. Quantitative evaluation reveals that indirect mapping can provide finer, smoother, and more efficient motion control than does direct mapping by operating in a continuous, rather than discrete, motor control space. It is also shown to outperform basis function neural networks. Furthermore, training its control parameters with recursive least squares enables faster convergence and better performance compared to gradient descent. The cooperation and competition of multiple selforganized EKMs allow a nonholonomic mobile robot to negotiate unforeseen, concave, closely spaced, and dynamic obstacles. Qualitative and quantitative comparisons with neural network ensembles employing weighted sum reveal that our method can achieve more sophisticated motion tasks even though the weightedsum ensemble approach also operates in continuous motor control space.
 Task Allocation via SelfOrganizing Swarm Coalitions in Distributed Mobile Sensor Network.
Kian Hsiang Low, Wee Kheng Leow^{} & Marcelo H. Ang, Jr.^{}
In Proceedings of the 19th National Conference on Artificial Intelligence (AAAI04), pages 2833, San Jose, CA, Jul 2529, 2004.
26.7% acceptance rate
Abstract. This paper presents a task allocation scheme via selforganizing swarm coalitions for distributed mobile sensor network coverage. Our approach uses the concepts of ant behavior to selfregulate the regional distributions of sensors in proportion to that of the moving targets to be tracked in a nonstationary environment. As a result, the adverse effects of task interference between robots are minimized and sensor network coverage is improved. Quantitative comparisons with other tracking strategies such as static sensor placement, potential fields, and auctionbased negotiation show that our approach can provide better coverage and greater flexibility to respond to environmental changes.
 Reactive, Distributed Layered Architecture for ResourceBounded MultiRobot Cooperation: Application to Mobile Sensor Network Coverage.
Kian Hsiang Low, Wee Kheng Leow^{} & Marcelo H. Ang, Jr.^{}
In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA'04), pages 37473752, New Orleans, LA, Apr 26  May 1, 2004.
Abstract. This paper describes a reactive, distributed layered architecture for cooperation of multiple resourcebounded robots, which is utilized in mobile sensor network coverage. In the upper layer, a dynamic task allocation scheme selforganizes the robot coalitions to track efficiently in separate regions. It uses the concepts of ant behavior to selfregulate the regional distributions of robots in proportion to that of the targets to be tracked in the changing environment. As a result, the adverse effects of task interference between robots are minimized and sensor network coverage is improved. In the lower layer, the robots use selforganizing neural networks to coordinate their target tracking within a region. Quantitative comparisons with other tracking strategies such as static sensor placements, potential fields, and auctionbased negotiation show that our approach can provide better coverage and greater flexibility in responding to environmental changes.
 ContinuousSpaced Action Selection for Single and MultiRobot Tasks Using Cooperative Extended Kohonen Maps.
Kian Hsiang Low, Wee Kheng Leow^{} & Marcelo H. Ang, Jr.^{}
In Proceedings of the IEEE International Conference on Networking, Sensing and Control (ICNSC'04)
(Invited Paper to Special Session on Visual Surveillance), pages 198203, Taipei, Taiwan, Mar 2123, 2004.
Abstract. Action selection is a central issue in the design of behaviorbased control architectures for autonomous mobile robots. This paper presents an action selection framework based on an assemblage of selforganizing neural networks called Cooperative Extended Kohonen Maps. This framework encapsulates two features that significantly enhance a robot's action selection capability: selforganization in the continuous state and action spaces to provide smooth, efficient and fine motion control; action selection via the cooperation and competition of Extended Kohonen Maps so that more complex motion tasks can be achieved. Qualitative and quantitative comparisons for both single and multirobot motion tasks show that our framework can provide better action selection than do action superposition methods.
 Action Selection for Single and MultiRobot Tasks Using Cooperative Extended Kohonen Maps.
Kian Hsiang Low, Wee Kheng Leow^{} & Marcelo H. Ang, Jr.^{}
In Proceedings of the 18th International Joint Conference on Artificial Intelligence (IJCAI03), pages 15051506, Acapulco, Mexico, Aug 915, 2003.
27.6% acceptance rate
Abstract. This paper presents an action selection framework based on an assemblage of selforganizing neural networks called Cooperative Extended Kohonen Maps. This framework encapsulates two features that significantly enhance a robot's action selection capability: selforganization in the continuous state and action spaces to provide smooth, efficient and fine motion control; action selection via the cooperation and competition of Extended Kohonen Maps to achieve more complex motion tasks. Qualitative and quantitative comparisons for single and multirobot tasks show our framework can provide better action selection than do potential fields method.
 Action Selection in Continuous State and Action Spaces by Cooperation and Competition of Extended Kohonen Maps.
Kian Hsiang Low, Wee Kheng Leow^{} & Marcelo H. Ang, Jr.^{}
In Proceedings of the
2nd International Joint Conference on Autonomous Agents and MultiAgent Systems (AAMAS03), pages 10561057, Melbourne, Australia, Jul 1418, 2003.
Abstract. This paper presents an action selection framework based on an assemblage of selforganizing neural networks called Cooperative Extended Kohonen Maps. This framework encapsulates two features that significantly enhance a robot's action selection capability: selforganization in the continuous state and action spaces to provide smooth, efficient and fine motion control; action selection via the cooperation and competition of Extended Kohonen Maps to achieve more complex motion tasks. Qualitative tests demonstrate the capability of our action selection method for both single and multirobot motion tasks.
 Enhancing the Reactive Capabilities of Integrated Planning and Control with Cooperative Extended Kohonen Maps.
Kian Hsiang Low, Wee Kheng Leow^{} & Marcelo H. Ang, Jr.^{}
In Proceedings of the
IEEE International Conference on Robotics and Automation (ICRA'03), pages 34283433, Taipei, Taiwan, May 1217, 2003.
Abstract. Despite the many significant advances made in robot motion research, few works have focused on the tight integration of highlevel deliberative planning with reactive control at the lowest level. In particular, the realtime performance of existing integrated planning and control architectures is still not optimal because the reactive control capabilities have not been fully realized. This paper aims to enhance the lowlevel reactive capabilities of integrated planning and control with Cooperative Extended Kohonen Maps for handling complex, unpredictable environments so that the workload of the highlevel planner can be consequently eased. The enhancements include fine, smooth motion control, execution of more complex motion tasks such as overcoming unforeseen concave obstacles and traversing between closely spaced obstacles, and asynchronous execution of behaviors.
 A Hybrid Mobile Robot Architecture with Integrated Planning and Control.
Kian Hsiang Low, Wee Kheng Leow^{} & Marcelo H. Ang, Jr.^{}
In Proceedings of the
1st International Joint Conference on Autonomous Agents and MultiAgent Systems (AAMAS02), pages 219226, Bologna, Italy, Jul 1519, 2002.
26% acceptance rate
Abstract. Research in the planning and control of mobile robots has received much attention in the past two decades. Two basic approaches have emerged from these research efforts: deliberative vs. reactive. These two approaches can be distinguished by their different usage of sensed data and global knowledge, speed of response, reasoning capability, and complexity of computation. Their strengths are complementary and their weaknesses can be mitigated by combining the two approaches in a hybrid architecture. This paper describes a method for goaldirected, collisionfree navigation in unpredictable environments that employs a behaviorbased hybrid architecture with asynchronously operating behavioral modules. It differs from existing hybrid architectures in two important ways: (1) the planning module produces a sequence of checkpoints instead of a conventional complete path, and (2) in addition to obstacle avoidance, the reactive module also performs target reaching under the control of a selforganizing neural network. The neural network is trained to perform fine, smooth motor control that moves the robot through the checkpoints. These two aspects facilitate a tight integration between highlevel planning and lowlevel control, which permits realtime performance and easy path modification even when the robot is en route to the goal position.
 Integrated Planning and Control of Mobile Robot with SelfOrganizing Neural Network.
Kian Hsiang Low, Wee Kheng Leow^{} & Marcelo H. Ang, Jr.^{}
In Proceedings of the
IEEE International Conference on Robotics and Automation (ICRA'02), pages 38703875, Washington, DC, May 1115, 2002.
Abstract. Despite the many significant advances made in robotics research, few works have focused on the tight integration of task planning and motion control. Most integration works involve the task planner providing discrete commands to the lowlevel controller, which performs kinematics and control computations to command the motor and joint actuators. This paper presents a framework of the integrated planning and control for mobile robot navigation. Unlike existing integrated approaches, it produces a sequence of checkpoints instead of a complete path at the planning level. At the motion control level, a neural network is trained to perform motor control that moves the robot from one checkpoint to the next. This method allows for a tight integration between highlevel planning and lowlevel control, which permits realtime performance and easy modification of motion path while the robot is enroute to the goal position.

TECHNICAL REPORTS
 Understanding and Improving Neural Architecture Search.
Yao Shu^{}.
Ph.D. Thesis, Department of Computer Science, National University of Singapore, Jan 2022.
Abstract.
Over the past decade, various famous deep neural network (DNN) architectures have been devised and have achieved superhuman performance for a wide range of tasks. Designing these neural networks, however, typically incurs substantial efforts from domain experts by trials and errors. Such human efforts gradually become unaffordable with an increasing demand for customizing DNNs for different tasks. To this end, neural architecture search (NAS) has been widely applied to automate the design of neural networks in recent years. In the literature, a number of NAS algorithms have been proposed, aiming to further improve the search efficiency and effectiveness of NAS, i.e., to reduce the search cost and improve the generalization performance of the selected architectures, respectively. Despite these advances, there are still certain essential aspects in NAS that have not been well investigated in the literature, which however may help us to understand and even further improve popular NAS algorithms.
Firstly, only a few efforts have been devoted to understanding the neural architectures selected by popular NAS algorithms in the literature. In the first work of this thesis, we take the first step of understanding popular NAS algorithms by answering the following questions: What types of architectures are selected by popular NAS algorithms and why they are selected? In particular, we reveal that existing NAS algorithms (e.g., DARTS, ENAS) tend to favor architectures with wide and shallow cell structures. These favorable architectures consistently achieve fast convergence and are consequently selected by NAS algorithms. Our empirical and theoretical studies further confirm that their fast convergence derives from their smooth loss landscape and accurate gradient information. Nonetheless, these architectures may not necessarily lead to better generalization performance than other candidate architectures in the same search space, and therefore further improvement is possible by revising existing NAS algorithms.
Secondly, standard NAS algorithms typically aim to select only a single neural architecture from the search spaces and thus have overlooked the capability of other candidate architectures in helping improve the performance of their final selected architecture. To this end, we present two novel sampling algorithms under our Neural Ensemble Search via Bayesian Sampling (NESBS) framework that can effectively and efficiently select a wellperforming ensemble of neural architectures from NAS search space. Compared with stateoftheart NAS algorithms and other wellknown ensemble search baselines, our NESBS algorithms are shown to be able to achieve improved performance in both classification and adversarial defense tasks on various benchmark datasets while incurring a comparable search cost to these NAS algorithms.
Thirdly, the search efficiency of popular NAS algorithms in the literature is severely limited by the need for model training during the search process. To overcome this limitation, we propose a novel NAS algorithm called NAS at Initialization (NASI) that exploits the capability of Neural Tangent Kernel (NTK) in being able to characterize the converged performance of candidate architectures at initialization, hence allowing model training to be completely avoided to boost the search efficiency. Besides the improved search efficiency, NASI also achieves competitive search effectiveness on various datasets like CIFAR10/100 and ImageNet. Further, NASI can guarantee the benefits of being label and dataagnostic under mild conditions, i.e., the provable transferability of architectures selected by our NASI over different datasets.
Finally, though recent NAS algorithms using trainingfree metrics are able to select wellperforming architectures in practice, the reason why trainingfree NAS using these metrics performs well and the answer to the question of how trainingfree NAS can be further boosted still have not been fully understood. To this end, we provide a unified theoretical analysis for gradientbased trainingfree NAS in this paper to understand why trainingfree metrics work well in practice. By exploiting these theoretical understandings, we then develop a novel NAS framework called Hybrid Neural Architecture Search (HNAS) that consistently improves trainingfree NAS in a principled way. Remarkably, HNAS can enjoy the advantages of both trainingfree (i.e., the superior search efficiency) and trainingbased NAS (i.e., the remarkable search effectiveness), which we have demonstrated through extensive experiments.
 Exploiting Gradient Information for Modern Machine Learning Problems.
Yizhou Chen^{}.
Ph.D. Thesis, Department of Computer Science, National University of Singapore, Jan 2022.
Abstract. Many deep learning achievements are attributed to the backpropagation (BP) algorithm, which exploits gradient information of the deep neural network (DNN) models: BP efficiently computes the gradient of the loss function with respect to the weights of a DNN for a batch of examples, and such gradient can be used by stochastic gradient descent to perform learning / optimization of the DNN model. Despite recent advances in deep learning like DNN training, there are still important scenarios where we can also use gradient to tackle optimization difficulty. In a broader aspect of deep learning rather than DNN training, a significant challenge faced by ML practitioners is thus whether we can design efficient algorithms to use the model gradient in the training / optimization in various deep learning scenarios. This thesis identifies four important scenarios and, for each of them, proposes a novel algorithm to utilize the gradient information for effective optimization that is both theoretically grounded and practically effective.
Firstly, the training process of a machine learning (ML) model may be subject to adversarial attacks from an attacker who attempts to undermine the test performance of the ML model by perturbing the training minibatches, and thus needs to be protected by a defender. Such a problem setting is referred to as trainingtime adversarial ML. We formulate it as a twoplayer game and propose a principled Recursive Reasoningbased TrainingTime adversarial ML (R2T2) framework to model this game. R2T2 models the reasoning process between the attacker and the defender and captures their bounded reasoning capabilities (due to bounded computational resources) through the recursive reasoning formalism. In particular, we associate a deeper level of recursive reasoning with the use of a higherorder gradient to derive the attack (defense) strategy, which naturally improves its performance while requiring greater computational resources. R2T2 can empirically achieve stateoftheart attack and defense performances on benchmark image datasets.
Secondly, probabilistic modeling with neural network architectures constitute a wellstudied and popular area of deep learning. In contrast to a frequentist approach which is easy to overfit to the available dataset and risk learning unwanted biasing in the dataset, Gaussian process (GP) models were introduced as a fully probabilistic substitute and is one of the dominant approaches in Bayesian learning. A multilayer deep Gaussian process (DGP) model is a hierarchical composition of GP models with a greater expressive power, and is more useful when dealing with complicated dataset. Exact DGP inference is intractable, and the approximation methods either yields a biased posterior belief (deterministic approximation by variational inference) or is computationally costly (stochastic approximation by Monte Carlo sampling). These difficulties have motivated our recent development of an implicit posterior variational inference (IPVI) framework for DGPs that can ideally recover an unbiased posterior belief and still preserve time efficiency. However, as a generator and a discriminator are integrated in each layer of the DGP, the training becomes unstable and is prone to optimization difficulties. To resolve such issues, we propose a novel gradientbridging architecture of the generator and discriminator for the DGP model, which uses the inducing inputs as the context, thus leads to faster training and more accurate predictions. Empirical evaluation shows that IPVI with our proposed architecture outperforms the stateoftheart methods for DGPs.
Thirdly, many widely adopted Bayesian metalearning frameworks model the uncertainty in the predictions with a set of particles or a variational distribution (of the
metaparameters), which does not allow latent task modeling.1 We present a novel
implicit processbased metalearning (IPML) algorithm that, in contrast to existing works, explicitly represents each task as a continuous latent vector and models its probabilistic belief within the highly expressive implicit processes (IP) framework. IP is a stochastic process with highly flexible implicit priors over functions, and is suitable as a Bayesian (meta) learning model for complicated datasets (e.g., when the priors are nonGaussian unlike in GP). We tackle the metatraining in IPML with a novel expectationmaximization algorithm based on the stochastic gradient Hamiltonian Monte Carlo sampling method. Our delicate design of the neural network architecture for metatraining in IPML allows competitive metalearning performance to be achieved. Unlike existing works, IPML offers the benefits of being amenable to the characterization of a principled distance measure between tasks using the maximum mean discrepancy, active task selection without needing the assumption of known task contexts, and synthetic task generation by modeling taskdependent input distributions. Empirical evaluation on benchmark datasets shows that IPML outperforms existing Bayesian metalearning algorithms.
Last but not least, in the problem of active task selection which involves selecting the most informative tasks for metalearning, we propose a novel active task selection criterion based on the mutual information between latent task vectors. Unfortunately, such a criterion scales poorly in the number of candidate tasks when optimized. To resolve this issue, we exploit the submodularity property of our new criterion for devising the first active task selection algorithm for metalearning with a nearoptimal performance guarantee. To further improve our efficiency, we propose an online variant of the Stein variational gradient descent to perform fast belief updates of the metaparameters via maintaining a set of forward (and backward) particles when learning (or unlearning) from each selected task. We empirically demonstrate the superior performance of our proposed algorithm on realworld datasets.
 SampleEfficient Automated Machine Learning with Bayesian Optimization.
Zhongxiang Dai^{}.
Ph.D. Thesis, Department of Computer Science, National University of Singapore, Jul 2021.
Abstract. Automated hyperparameter optimization of machine learning (ML) models, referred to as AutoML, has been a challenging problem for practitioners, mainly due to the high computational cost of training modern ML models and the lack of gradient information with respect to the model hyperparameters. To this end, the blackbox optimization method of Bayesian optimization (BO) has become a prominent method for optimizing the hyperparameters of ML models, which can be attributed to its impressive sample efficiency and theoretical convergence guarantee. Despite recent advances, there are still important scenarios where we can further improve the sample efficiency of BO for AutoML by exploiting naturally available auxiliary information, or extend the applicability of BO to other ML tasks. This thesis identifies five such important scenarios and, for each of them, proposes a novel BO algorithm that is both theoretically grounded and practically effective.
Firstly, many ML models require an iterative training process, which requires every hyperparameter evaluation during BO to run for a certain number of training epochs. As a result, the auxiliary observations from intermediate training epochs can be exploited to earlystop the evaluations of those unpromising hyperparameter configurations to save resource. We propose the BO with Bayesian optimal stopping (BOBOS) algorithm, which incorporates BOS into BO in order to improve the epoch efficiency of BO using a principled optimal stopping mechanism. BOBOS preserves the asymptotic noregret property of BO with our specified setting of BOS parameters which is amenable to an elegant interpretation in terms of the explorationexploitation tradeoff, and performs competitively in a number of AutoML experiments.
Secondly, the widely celebrated federated learning (FL) setting requires firstorder optimization techniques, and is hence unable to handle zerothorder optimization tasks such as hyperparameter optimization. We extend BO into the FL setting (FBO) and derive the federated Thompson sampling (FTS) algorithm, to improve the efficiency of BO in the FL setting by employing auxiliary information from other agents. FTS tackles a number of major challenges faced by FBO in a principled way: FTS uses random Fourier features approximation to derive the parameters to be communicated in order to avoid sharing the raw data, adopts the Thompson sampling algorithm which reduces the number of parameters to be exchanged, and is robust against heterogeneous agents due to a robust theoretical convergence guarantee.
Thirdly, the abovementioned FTS algorithm, unfortunately, is not equipped with a rigorous privacy guarantee, which is an important consideration in FL. To this end, we integrate differential privacy (DP) into FTS through a general framework for adding DP to iterative algorithms. Moreover, we leverage the ability of this general DP framework to handle different parameter vectors, as well as the technique of local modeling for BO, to further improve the utility of our algorithm through distributed exploration (DE). The resulting DPFTSDE algorithm is able to improve an agent’s sample efficiency by exploiting auxiliary information from other agents, while rigorously hiding its participation in the algorithm. DPFTSDE is amenable to a number of interesting theoretical insights regarding the privacyutility tradeoff, and achieves competitive utilities with strong privacy guarantees in realworld experiments.
Fourthly, when BO is used for hyperparameter optimization using a dataset, we often have access to previous completed hyperparameter optimization tasks using other potentially related datasets. This prompts the question as to whether we can leverage these previous completed tasks to improve the efficiency of the current BO task through metalearning, while ensuring its robustness against dissimilar tasks. We introduce a scalable, principled and robust metaBO algorithm called robust metaGaussian processupper confidence bound (RMGPUCB). We show that RMGPUCB is asymptotically noregret even when all previous tasks are dissimilar to the current task, and is amenable to a principled method to learn the weights assigned to the individual previous tasks through regret minimization via online learning. RMGPUCB achieves effective performances in a wide range of realworld experiments.
Lastly, many ML tasks such as adversarial ML can be modeled as repeated games between boundedly rational, selfinterested agents with unknown, complex, and costlytoevaluate payoff functions. We introduce a recursive reasoning formalism of BO, called Recursive ReasoningBased BO (R2B2), which extends the applicability of BO to provide efficient strategies for players in this type of game. Under certain conditions, using R2B2 to reason at one level higher than the other agents achieves faster asymptotic convergence to no regret than without using recursive reasoning. R2B2 performs effectively in practice in adversarial ML and multiagent reinforcement learning experiments.
 Automated Machine Learning: New Advances on Bayesian Optimization.
Dmitrii Kharkovskii^{}.
Ph.D. Thesis, Department of Computer Science, National University of Singapore, Dec 2020.
Abstract. Recent advances in Bayesian optimization (BO) have delivered a promising suite of tools for optimizing an unknown expensive to evaluate blackbox objective function with a finite budget of evaluations. A significant advantage of BO is its general formulation: BO can be utilized to optimize any blackbox objective function. As a result, BO has been applied in a wide range of applications such as automated machine learning, robotics or environmental monitoring, among others. Furthermore, its general formulation makes BO attractive for deployment in new applications. However, potential new applications can have additional requirements not satisfied by the classical BO setting. In this thesis, we aim to address some of these requirements in order to scale up BO technology for the practical use in new realworld applications.
Firstly, this thesis tackles the problem of data privacy, which is not addressed by the standard setting of BO. Specifically, we consider the outsourced setting where the entity holding the dataset and the entity performing BO are represented by different parties, and the dataset cannot be released nonprivately. For example, a hospital holds a dataset of sensitive medical records and outsources the BO task on this dataset to an industrial AI company. We present the privateoutsourcedGaussian processupper confidence bound (POGPUCB) algorithm, which is the first algorithm for privacypreserving BO in the outsourced setting with a provable performance guarantee. The key idea of our approach is to make the BO performance of our algorithm similar to that of nonprivate GPUCB run using the original dataset, which is achieved by using a random projectionbased transformation that preserves both privacy and the pairwise distances between inputs. Our main theoretical contribution is to show that a regret bound similar to that of the standard GPUCB algorithm can be established for our POGPUCB algorithm. We empirically evaluate the performance of our algorithm with synthetic and realworld datasets.
Secondly, we consider applications of BO for hotspot sampling in spatially varying phenomena. For such applications, we exploit the structure of the spatially varying phenomenon in order to increase the BO lookahead and, as a result, improve the performance of the algorithm and make it more suitable for practical use in realworld scenarios. To do this, we present a principled multistaged Bayesian sequential decision algorithm for nonmyopic adaptive BO that, in particular, exploits macroactions for scaling up to a further lookahead to match up to a larger available budget. To achieve this, we first generalize GPUCB to a new acquisition function defined with respect to a nonmyopic adaptive macroaction policy, which, unfortunately, is intractable to be optimized exactly due to an uncountable set of candidate outputs. The key novel contribution of our work here is to show that it is in fact possible to solve for a nonmyopic adaptive εBayesoptimal macroaction BO (εMacroBO) policy given an arbitrary userspecified loss bound ε via stochastic sampling in each planning stage which requires only a polynomial number of samples in the length of macroactions. To perform nonmyopic adaptive BO in real time, we then propose an asymptotically optimal anytime variant of our εMacroBO algorithm with a performance guarantee. Empirical evaluation on synthetic and realworld datasets shows that our proposed approach outperforms existing stateoftheart algorithms.
Finally, this thesis proposes a blackbox attack for adversarial machine learning based on BO. Since the dimension of the inputs in adversarial learning is usually too high for applying BO directly, our proposed attack applies dimensionality reduction and searches for an adversarial perturbation in a lowdimensional latent space. The key idea of our approach is to automate both the selection of the latent space dimen sion and the search of the adversarial perturbation in the selected latent space by using BO. Additionally, we use Bayesian optimal stopping to boost the query efficiency of our attack. Performance evaluation using image classification datasets shows that our proposed method outperforms the stateoftheart blackbox adversarial attacks.
 New Advances in Bayesian Inference for Gaussian Process and Deep Gaussian Process Models.
Haibin Yu^{}.
Ph.D. Thesis, Department of Computer Science, National University of Singapore, May 2020.
Abstract.
Machine learning is the study of letting computers learn to perform a specific task in a datadriven manner. In particular, Bayesian machine learning has attracted enor mous attention mainly due to their ability to provide uncertainty estimates following Bayesian inference. This thesis focuses on Gaussian processes (GPs), a rich class of Bayesian nonparametric models for performing Bayesian machine learning with formal measures of predictive uncertainty.
However, the applicability of GP in large datasets and in hierarchical composition of GPs is severely limited by computational issues and intractabilities. Therefore, it is crucial to develop accurate and efficient inference algorithms to address these challenges. To this end, this thesis aims at proposing a series of novel approximate Bayesian inference methods for a wide variety of GP models, which unifies the previous literatures, significantly extends them and hopefully lays the foundation for future inference methods.
To start with, this thesis presents a unifying perspective of existing inducing variablesbased GP models, sparse GP (SGP) models and variational inference for SGP models (VSGP). Then, to further mitigate the issue of overfitting during optimization, we present a novel variational inference framework for deriving a family of Bayesian SGP regression models, referred to as variational Bayesian SGP (VBSGP) regression models.
Next, taking into account the fact that the expressiveness of GP and SGP depends heavily on the design of the kernel function, we further extend the expressive power of GP by introducing Deep GP (DGP), which is a hierarchical composition of GP models. Unfortunately, exact inference in DGP is intractable, which has motivated the recent development of deterministic and stochastic approximation methods. However, the deterministic approximation methods yield a biased posterior belief while the stochastic one is computationally costly. In this regard, we present the implicit posterior variational inference (IPVI) framework for DGPs that can ideally recover an unbiased posterior belief and still preserve time efficiency. Inspired by generative adversarial networks, our IPVI framework casts the DGP inference problem as a two player game in which a Nash equilibrium, interestingly, coincides with an unbiased posterior belief.
We hope this thesis at least provides additional confidence and clarity for researchers who are devoting themselves to Bayesian nonparametric models, Gaussian process models in particular. Moreover, we also wish this thesis to offer inspirations for future works, and some thoughts that could be useful for future solutions.
 DataEfficient Machine Learning with Multiple Output Types and High Input Dimensions.
Yehong Zhang^{}.
Ph.D. Thesis, Department of Computer Science, National University of Singapore, Dec 2017.
Abstract.
Recent research works in machine learning (ML) have focused on learning some target variables of interest to achieve competitive (or stateoftheart) predictive performance in less time but without requiring large quantities of data, which is known as dataefficient ML. This thesis focuses on two important dataefficient ML approaches: active learning (AL) and Bayesian optimization (BO) which, instead of learning passively from a given small set of data, need to select and gather the most informative observations for learning the target variables of interest more accurately given some budget constraints. To advance the stateoftheart of dataefficient ML, novel generalizations of AL and BO algorithms are proposed in this thesis for addressing the issues arising from multiple output types and high input dimensions which are the practical settings in many realworld applications.
In particular, this thesis aims to (a) exploit the auxiliary types of outputs which usually coexist and correlate well with the target output types, and more importantly, are less noisy and/or less tedious to sample for improving the learning performance of the target output type in both AL and BO algorithms and (b) scale up the stateoftheart BO algorithm to high input dimensions. To achieve this, the specific data with multiple output types or high input dimensions is represented using some form of Gaussian process (GP)based probabilistic regression models which allow the predictive uncertainty of the outputs to be formally quantified and consequently exploited for developing efficient AL and BO algorithms.
To achieve above objectives, an AL algorithm of multioutput GP (MOGP) is first developed for minimizing the predictive uncertainty (i.e., posterior joint entropy) of the target output type. In contrast to existing works, our AL problems involve selecting not just the most informative sampling inputs to be observed but also the types of outputs at each selected input for improving the learning performance of only the target output type given a sampling budget. Unfortunately, such an entropy criterion scales poorly in the numbers of candidate sampling inputs and selected observations when optimized. To resolve this issue, we exploit a structure common to sparse MOGP models for deriving a novel AL criterion. Furthermore, we exploit a relaxed form of submodularity property of our new criterion for devising a polynomialtime approximation algorithm that guarantees a constantfactor approximation of that achieved by the optimal set of selected observations. Empirical evaluation on realworld datasets shows that our proposed approach outperforms existing algorithms for AL of MOGP and singleoutput GP models.
Secondly, to boost the BO performance by exploiting the cheaper and/or less noisy observations of some auxiliary functions with varying fidelities, we proposed a novel generalization of predictive entropy search (PES) for multifidelity BO called multifidelity PES (MFPES). In contrast to existing multifidelity BO algorithms, our proposed MFPES algorithm can naturally trade off between exploitation vs. exploration over the target and auxiliary functions with varying fidelities without needing to manually tune any such parameters. To achieve this, we model the unknown target and auxiliary functions jointly as a convolved MOGP (CMOGP) whose convolutional structure is exploited to formally characterize the fidelity of each auxiliary function through its crosscorrelation with the target function. Although the exact acquisition function of MFPES cannot be computed in closed form, we show that it is in fact possible to derive an efficient approximation of MFPES via a novel multioutput random features approximation of the CMOGP model whose crosscorrelation (i.e., multifidelity) structure between the target and auxiliary functions can be exploited for improving the belief of the global target maximizer using the observations from evaluating these functions. Practical constraints are proposed to relate the global target maximizer to that of auxiliary functions. Empirical evaluation on synthetic and realworld experiments shows that MFPES outperforms the stateoftheart multifidelity BO algorithms.
Lastly, to improve the BO performance in realworld applications with high input
dimensions (e.g., computer vision, biology), we generalize PES for highdimensional BO by exploiting an additive structure of the target function. New practical constraints are proposed and approximated efficiently such that the proposed acquisition function of additive PES (addPES) can be optimized independently for each local and lowdimensional input component. The empirical results show that our addPES considerably improves the performance of the stateoftheart highdimensional BO algorithms by using a simple and common setting for optimizing different tested functions with varying input dimensions, which makes it a superior alternative to existing highdimensional BO algorithms.
 Exploiting Decentralized MultiAgent Coordination for LargeScale Machine Learning Problems.
Ruofei Ouyang^{}.
Ph.D. Thesis, Department of Computer Science, National University of Singapore, Dec 2016.
Abstract.
Nowadays, the scale of machine learning problems becomes much larger than before. It raises a huge demand in distributed perception and distributed computation. A multiagent system provides exceptional scalability for problems like active sensing and data fusion. However, many rich characteristics of largescale machine learning problems have not been addressed yet such as large input domain, nonstationarity, and high dimensionality. This thesis identifies the challenges related to these characteristics from multiagent perspective. By exploiting the correlation structure of data in largescale problems, we propose multiagent coordination schemes that can improve the scalability of the machine learning models while preserving the computation accuracy. To elaborate, the machine learning problems we are solving with multiagent coordination techniques are:
 Gaussian process regression. To perform distributed regression on a largescale environmental phenomenon, data compression is often required due to the communication costs. Currently, decentralized data fusion methods encapsulate the data into local summaries based on a fixed support set. However in a largescale field, this fixed support set, acting as a centralized component in the decentralized system, cannot approximate the correlation structure of the entire phenomenon well. It leads to evident losses in data summarization. Consequently, the regression performance will be significantly reduced.
In order to approximate the correlation structure accurately, we propose an agentcentric support set to allow every agent in the data fusion system to choose a possibly different support set and dynamically switch to another one during execution for encapsulating its own data into a local summary which, perhaps surprisingly, can still be assimilated with the other agents’ local summaries into a globally consistent summary. Together with an information sharing mechanism we designed, the new decentralized data fusion methods with agentcentric support set can be applied to regression problems on a much larger environmental phenomenon with high performance.
 Active learning. In the context of environmental sensing, active learning/active sensing is a process of taking observations to minimize the uncertainty in an environmental field. The uncertainty is quantified based on the correlation structure of the phenomenon which is traditionally assumed to be stationary for computational sake. In a largescale environmental field, this stationary assumption is often violated. Therefore, existing active sensing algorithms perform suboptimally for a nonstationary environmental phenomenon.
To the best of our knowledge, our decentralized multirobot active sensing (DECMAS) algorithm is the first work to address nonstationarity issue in the context of active sensing. The uncertainty in the phenomenon is quantified based on the nonstationary correlation structure estimated by Dirichlet process mixture of Gaussian processes. Further, our DECMAS algorithm can efficiently coordinate the exploration of multiple robots to automatically tradeoff between learning the unknown, nonstationary correlation structure and minimizing the uncertainty of the environmental phenomenon. It enables multiagent active sensing techniques to be applied to a largescale nonstationary environmental phenomenon.
 Bayesian optimization. Optimizing an unknown objective function is challenging for traditional optimization methods. Alternatively, in this situation, people use Bayesian optimization which is a modern optimization technique that can optimize a function by only utilizing the observation information (input and output values) collected through simulations. When the input dimension of the function is low, a few simulated observations can generate good result already. However, for high dimensional function, a huge number of observations are required which is impractical when the simulation consumes lots of time and resources.
Fortunately, many high dimensional problems have sparse correlation structure. Our ANOVADCOP work can decompose the correlation structure in the original highdimensional problem into many correlation structures of subsets of dimensions based on ANOVA kernel function. It significantly reduces the size of input space into a collection of lowerdimensional subspaces. Additionally, we reformulate the Bayesian optimization problem as a decentralized constrained optimization problem (DCOP) that can be efficiently solved by multiagent coordination techniques so that it can scale up to problems with hundreds of dimensions.
 New Advances on Bayesian and DecisionTheoretic Approaches for Interactive Machine Learning.
Trong Nghia Hoang^{}.
Ph.D. Thesis, Department of Computer Science, National University of Singapore, Feb 2015.
Abstract.
The explorationexploitation tradeoff is a fundamental dilemma in many interactive learning scenarios which include both aspects of reinforcement learning (RL) and active learning (AL): An autonomous agent, situated in an unknown environment, has to actively extract knowledge from the environment by taking actions (or conducting experiments) based on its previously collected information to make accurate predictions or to optimize some utility functions. Thus, to make the most effective use of their resourceconstrained budget (e.g., processing time, experimentation cost), the agent must choose carefully between (a) exploiting options (e.g., actions, experiments) which are recommended by its current, possibly incomplete model of the environment, and (b) exploring the other ostensibly suboptimal choices to gather more information.
For example, an RL agent has to face a dilemma between (a) exploiting the mostrewarding action according to the current statistical model of the environment at the risk of running into catastrophic situations if the model is not accurate, and (b) exploring a suboptimal action to gather more information so as to improve the model's accuracy at the potential price of losing the shortterm reward. Similarly, an AL algorithm/agent has to consider between (a) conducting the most informative experiments according to its current estimation of the environment model's parameters (i.e., exploitation), and (b) running experiments that help improving the estimation accuracy of these parameters (i.e., exploration).
More often, learning strategies that ignore exploration will likely exhibit suboptimal performance due to their imperfect knowledge while, conversely, those that entirely focus on exploration might suffer the cost of learning without benefitting from it. Therefore, a good explorationexploitation tradeoff is critical to the success of those interactive learning agents: In order to perform well, they must strike the right balance between these two conflicting objectives. Unfortunately, while this tradeoff has been wellrecognized since the early days of RL, the studies of explorationexploitation have mostly been developed for theoretical settings in the respective field of RL and, perhaps surprisingly, glossed over in the existing AL literature. From a practical point of view, we see three limiting factors:
 Previous works addressing the explorationexploitation tradeoff in RL have largely focused on simple choices of the environment model and consequently, are not practical enough to accommodate realworld applications that have far more complicated environment structures. In fact, we find that most recent advances in Bayesian reinforcement learning (BRL) have only been able to analytically trade off between exploration and exploitation under a simple choice of models such as FlatDirichletMultinomial (FDM) whose independence and modeling assumptions do not hold for many realworld applications.
 Nearly all of the notable works in the AL literature primarily advocate the use of greedy/myopic algorithms whose rates of convergence (i.e., the number of experiments required by the learning algorithm to achieve a desired performance in the worst case) are provably minimax optimal for simple classes of learning tasks (e.g., threshold learning). While these results have greatly ad vanced our understanding about the limit of myopic AL in worstcase scenarios, significantly less is presently known about whether it is possible to devise nonmyopic AL strategies which optimize the explorationexploitation tradeoff to achieve the best expected performance in budgeted learning scenarios.
 The issue of scalability of the existing predictive models (e.g., Gaussian processes) used in AL has generally been underrated since the majority of literature considers smallscale environments which only consist of a few thousand candidate experiments to be selected by singlemode AL algorithms one at a time prior to retraining the model. In contrast, largescale environments usually have a massive set of million candidate experiments among which tens or hundreds of thousands should be actively selected for learning. For such dataintensive problems, it is often more costeffective to consider batchmode AL algorithms which select and conduct multiple experiments in parallel at each stage to collect observations in batch. Retraining the predictive model after incorporating each batch of observations then becomes a computational bottleneck as the collected dataset at each stage quickly grows up to tens or even hundreds of thousand data points.
This thesis outlines some recent progresses that we have been able to make while working toward satisfactory answers to the above challenges, along with practical algorithms that achieve them:
 In particular, in order to put BRL into practice for more complicated and practical problems, we propose a novel framework called Interactive Bayesian Reinforcement Learning (IBRL) to integrate the general class of parametric models and model priors, thus allowing the practitioners' domain knowledge to be exploited to produce a finegrained and compact representation of the environment as often required in many realworld applications. Interestingly, we show how the nonmyopic Bayesoptimal policy can be derived analytically by solving IBRL exactly and propose an approximation algorithm to compute it efficiently in polynomial time. Our empirical studies show that the proposed approach performs competitively with the existing stateoftheart algorithms.
 Then, to establish a theoretical foundation for the explorationexploitation tradeoff in singlemode active learning scenarios with resourceconstrained budgets, we present a novel ϵBayesoptimal DecisionTheoretic Active Learning (ϵBAL) framework which advocates the use of differential entropy as a performance measure and consequently, derives a learning policy that can approximate the optimal expected performance arbitrarily closely (i.e., within an arbitrary loss bound ϵ). To meet the realtime requirement in timecritical applications, we then propose an asymptotically ϵoptimal, branchandbound anytime algorithm based on ϵBAL with performance guarantees. In practice, we empirically demonstrate with both synthetic and realworld datasets that the proposed approach outperforms the stateoftheart algorithms in budgeted scenarios.
 Lastly, to facilitate the future developments of largescale, nonmyopic AL applications, we further introduce a highly scalable family of anytime predictive models for AL which provably converge toward a wellknown class of sparse Gaussian processes (SGPs). Unlike the existing predictive models of AL which cannot be updated incrementally and are only capable of processing middlesized datasets (i.e., a few thousands of data points), our proposed models can process massive datasets in an anytime fashion, thus providing a principled tradeoff between the processing time and the predictive accuracy. The efficiency of our framework is then demonstrated empirically on a variety of largescale realworld datasets which contains hundreds of thousand data points.
 Gaussian ProcessBased Decentralized Data Fusion and Active Sensing Agents: Towards LargeScale Modeling and Prediction of Spatiotemporal Traffic Phenomena.
Jie Chen^{}.
Ph.D. Thesis, Department of Computer Science, National University of Singapore, Dec 2013.
Abstract.
Knowing and understanding the environmental phenomena is important to many real world applications. This thesis is devoted to study largescale modeling and prediction of spatiotemporal environmental phenomena (i.e., urban traffic phenomena). Towards this goal, our proposed approaches rely on a class of Bayesian nonparametric models: Gaussian processes (GP).
To accurately model spatiotemporal urban traffic phenomena in real world situation, a novel relational GP taking into account both the road segment features and road network topology information is proposed to model real world traffic conditions over road network. Additionally, a GP variant called logGaussian process (lGP) is exploited to model an urban mobility demand pattern which contains skewness and extremity in demand measurements.
To achieve efficient and scalable urban traffic phenomenon prediction given a large phenomenon data, we propose three novel parallel GPs: parallel partially independent training conditional (pPITC), parallel partially independent conditional(pPIC) and parallel incomplete Cholesky factorization (pICF)based approximations of GP model, which can distribute their computational load into a cluster of parallel/multicore machines, thereby achieving time efficiency. The predictive performances of such parallel GPs are theoretically guaranteed to be equivalent to that of some centralized approaches to approximate full/exact GP regression. The proposed parallel GPs are implemented using the message passing interface (MPI) framework and tested on two large real world datasets. The theoretical and empirical results show that our parallel GPs achieve significantly better time efficiency and scalability than that of full GP, while achieving comparable accuracy. They also achieve fine speedup performance that is the ratio of time required by the parallel algorithms and their centralized counterparts.
To exploit active mobile sensors to perform decentralized perception of the spatiotemporal urban traffic phenomenon, we propose a decentralized algorithm framework: Gaussian processbased decentralized data fusion and active sensing (D2FAS) which is composed of a decentralized data fusion (DDF) component and a decentralized active sensing (DAS) component. The DDF component includes a novel Gaussian processbased decentralized data fusion (GPDDF) algorithm that can achieve remarkably efficient and scalable prediction of phenomenon and a novel Gaussian processbased decentralized data fusion with local augmentation (GPDDF+) algorithm that can achieve better predictive accuracy while preserving time efficiency of GPDDF. The predictive performances of both GPDDF and GPDDF+ are theoretically guaranteed to be equivalent to that of some sophisticated centralized sparse approximations of exact/full GP. For the DAS component, we propose a novel partially decentralized active sensing (PDAS) algorithm that exploits property in correlation structure of GPDDF to enable mobile sensors cooperatively gathering traffic phenomenon data along a nearoptimal joint walk with theoretical guarantee, and a fully decentralized active sensing (FDAS) algorithm that guides each mobile sensor gather phenomenon data along its locally optimal walk.
Lastly, to justify the practicality of the D2FAS framework, we develop and test D2FAS algorithms running with active mobile sensors on real world datasets for monitoring traffic conditions and sensing/servicing urban mobility demands. Theoretical and empirical results show that the proposed algorithms are significantly more timeefficient, more scalable in the size of data and in the number of sensors than the stateoftheart centralized approaches, while achieving comparable predictive accuracy.
 A DecisionTheoretic Approach for Controlling and Coordinating Multiple Active Cameras in Surveillance.
Prabhu Natarajan^{}.
Ph.D. Thesis, Department of Computer Science, National University of Singapore, Dec 2013.
Abstract.
The use of active cameras in surveillance is becoming increasingly popular as they try to meet the demands of capturing highresolution images/videos of targets in surveillance for face recognition, target identification, forensic video analysis, etc. These active cameras are endowed with pan, tilt, and zoom capabilities, which can be exploited to provide highquality surveillance. In order to achieve effective, realtime surveillance, an efficient collaborative mechanism is needed to control and coordinate these cameras' actions, which is the focus of this thesis. The central problem in surveillance is to monitor a set of targets with guaranteed image resolution. Controlling and coordinating multiple active cameras to achieve this surveillance task is nontrivial and challenging because: (a) presence of inherent uncertainties in the surveillance environment (targets motion, location, and noisy camera observation); (b) there exists a nontrivial tradeoff between number of targets and the resolution of observing these targets; and (c) more importantly, the coordination framework should be scalable with increasing number of targets and cameras.
In this thesis, we formulate a novel decisiontheoretic multiagent planning approach for controlling and coordinating multiple active cameras in surveillance. Our decisiontheoretic approach offers advantages of (a) accounting the uncertainties using probabilistic models; (b) the nontrivial tradeoff is addressed by coordinating the active cameras' actions to maximize the number of targets with guaranteed resolution; and (c) the scalability in number of targets and cameras is achieved by exploiting the structures and properties that are present in our surveillance problem. We focus on two novel problems in active camera surveillance: (a) maximizing observations of multiple targets (MOMT), i.e., maximizing the number of targets observed in active cameras with guaranteed image resolution; and (b) improving fairness in observation of multiple targets (FOMT), i.e., no target is "starved" of observation by active cameras for long duration of time.
We propose two formal decisiontheoretic frameworks (a) Markov Decision Process (MDP) and (b) Partially Observable Markov Decision Process (POMDP) frameworks for coordinating active cameras in surveillance. MDP framework controls active cameras in fully observable surveillance environments where the active cameras are supported by one or more wideview static/fixed cameras to observe the entire surveillance environment at lowresolution. POMDP framework controls active cameras in partially observable surveillance environments where it is impractical to observe the entire surveillance environment using static/fixed cameras due to occlusions caused by physical infrastructures. Hence the POMDP framework do not have a complete view of the surveillance environment.
Specifically, we propose (a) MDP frameworks to solve MOMT problem and FOMT problem in fully observable surveillance environment; and (b) POMDP framework to solve MOMT problem in partially observable surveillance environment. As proven analytically, our MDP and POMDP frameworks incurs time that is linear in number of targets to be observed during surveillance. We have used maxplus algorithm with our MDP framework to improve its scalability in number of cameras for MOMT problem. Empirical evaluation through simulations in realistic surveillance environment reveals that our proposed approach can achieve highquality surveillance in real time. We also demonstrate our pro posed approach with real Axis 214 PTZ cameras to show the practicality of our approach in real world surveillance. Both the simulations and real camera experiments show that our decisiontheoretic approach can control and coordinate active cameras efficiently and hence contributes significantly towards improving the active camera surveillance research.
 InformationTheoretic MultiRobot Path Planning.
Nannan Cao^{}.
M.Sc. Thesis, Department of Computer Science, National University of Singapore, Sep 2012.
Abstract.
Research in environmental sensing and monitoring is especially important in supporting environmental sustainability efforts worldwide, and has recently attracted significant attention and interest. A key direction of this research lies in modeling and predicting the spatiotemporally varying environmental phenomena. One approach is to use a team of robots to sample the area and model the measurement values at unobserved points. For smoothly varying and hotspot fields, there is some work which has been done to model the fields well. However, there is still a class of common environmental fields called anisotropic fields in which the spatial phenomena are highly correlated along one direction and less correlated along the perpendicular direction. We exploit the environmental structure to improve the sampling performance and time efficiency of planning for anisotropic fields.
In this thesis, we cast the planning problem into a stagewise decisiontheoretic problem. we adopt Gaussian Process to model spatial phenomena. Maximum entropy criterion and maximum mutual information criterion are used to measure the informativeness of the observation paths. It is found that for many GPs, correlation of two points exponentially decreases with the distance between the two points. With this property, for maximum entropy criterion, we propose a polynomialtime approximation algorithm, MEPP, to find the maximum entropy paths. We also provide a theoretical performance guarantee for this algorithm. For maximum mutual information criterion, we propose another polynomialtime approximation algorithm, M2IPP. Similar to the MEPP, a performance guarantee is also provided for this algorithm. We demonstrate the performance advantages of our algorithms on two real data sets. To get lower prediction error, three priciples have also been proposed to select the criterion for different environmental fields.
 MultiRobot Adaptive Exploration and Mapping for Environmental Sensing Applications.
Kian Hsiang Low.
Ph.D. Thesis, Technical Report CMUECE2009024, Department of Electrical and Computer Engineering, Carnegie Mellon University, Pittsburgh, PA, Aug 2009.
Abstract.
Recent research in robot exploration and mapping has focused on sampling hotspot fields, which often arise in environmental and ecological sensing applications. Such a hotspot field is characterized by continuous, positively skewed, spatially correlated measurements with the hotspots exhibiting extreme measurements and much higher spatial variability than the rest of the field.
To map a hotspot field of the above characterization, we assume that it is realized from nonparametric probabilistic models such as the Gaussian and logGaussian processes (respectively, GP and lGP), which can provide formal measures of map uncertainty. To learn a hotspot field map, the exploration strategy of a robot team then has to plan resourceconstrained observation paths that minimize the uncertainty of a spatial model of the hotspot field. This exploration problem is formalized in a sequential decisiontheoretic planning under uncertainty framework called the multirobot adaptive sampling problem (MASP). So, MASP can be viewed as a sequential, nonmyopic version of active learning. In contrast to finitestate Markov decision problems, MASP adopts a more complex but realistic continuousstate, nonMarkovian problem structure so that its induced exploration policy can be informed by the complete history of continuous, spatially correlated observations for selecting paths. It is unique in unifying formulations of nonmyopic exploration problems along the entire adaptivity spectrum, thus subsuming existing nonadaptive formulations and allowing the performance advantage of a more adaptive policy to be theoretically realized. Through MASP, it is demonstrated that a more adaptive strategy can exploit clustering phenomena in a hotspot field to produce lower expected map uncertainty. By measuring map uncertainty using the meansquared error criterion, a MASPbased exploration strategy consequently plans adaptive observation paths that minimize the expected posterior map error or equivalently, maximize the expected map error reduction.
The time complexity of solving MASP (approximately) depends on the map resolution, which limits its practical use in largescale, highresolution exploration and mapping. This computational difficulty is alleviated through an informationtheoretic approach to MASP (iMASP), which measures map uncertainty based on the entropy criterion instead. As a result, an iMASPbased exploration strategy plans adaptive observation paths that minimize the expected posterior map entropy or equivalently, maximize the expected entropy of observation paths. Unlike MASP, reformulating the costminimizing iMASP as a rewardmaximizing dual problem causes its time complexity of being solved approximately to be independent of the map resolution and less sensitive to larger robot team size as demonstrated both analytically and empirically. Furthermore, this rewardmaximizing dual transforms the widelyused nonadaptive maximum entropy sampling problem into a novel adaptive variant, thus improving the performance of the induced exploration policy.
One advantage stemming from the rewardmaximizing dual formulations of MASP and iMASP is that they allow observation selection properties of the induced exploration policies to be realized for sampling the hotspot field. These properties include adaptivity, hotspot sampling, and widearea coverage. We show that existing GPbased exploration strategies may not explore and map the hotspot field well with the selected observations because they are nonadaptive and perform only widearea coverage. In contrast, the lGPbased exploration policies can learn a highquality hotspot field map because they are adaptive and perform both widearea coverage and hotspot sampling.
The other advantage is that even though MASP and iMASP are nontrivial to solve due to their continuous state components, the convexity of their rewardmaximizing duals can be exploited to derive, in a computationally tractable manner, discretestate monotonebounding approximations and subsequently, approximately optimal exploration policies with theoretical performance guarantees. Anytime algorithms based on approximate MASP and iMASP are then proposed to alleviate the computational difficulty that arises from their nonMarkovian structure.
It is of practical interest to be able to quantitatively characterize the "hotspotness" of an environmental field. We propose a novel "hotspotness" index, which is defined in terms of the spatial correlation properties of the hotspot field. As a result, this index can be related to the intensity, size, and diffuseness of the hotspots in the field.
We also investigate how the spatial correlation properties of the hotspot field affect the performance advantage of adaptivity. In particular, we derive sufficient and necessary conditions of the spatial correlation properties for adaptive exploration to yield no performance advantage.
Lastly, we develop computationally efficient approximately optimal exploration strategies for sampling the GP by assuming the Markov property in iMASP planning. We provide theoretical guarantees on the performance of the Markovbased policies, which improve with decreasing spatial correlation. We evaluate empirically the effects of varying spatial correlations on the mapping performance of the Markovbased policies as well as whether these Markovbased path planners are timeefficient for the transect sampling task.
Through the abovementioned work, this thesis establishes the following two claims: (1) adaptive, nonmyopic exploration strategies can exploit clustering phenomena to plan observation paths that produce lower map uncertainty than nonadaptive, greedy methods; and (2) Markovbased exploration strategies can exploit small spatial correlation to plan observation paths which achieve map uncertainty comparable to that of nonMarkovian policies using significantly less planning time.
 Adaptive Sampling for MultiRobot Wide Area Prospecting.
Kian Hsiang Low, Geoffrey J. Gordon, John M. Dolan, and Pradeep Khosla.
In Technical Report CMURITR0551, Robotics Institute, Carnegie Mellon University, Pittsburgh, PA, Oct 2005.
Abstract. Prospecting for in situ mineral resources is essential for establishing settlements on the Moon and Mars. To reduce human effort and risk, it is desirable to build robotic systems to perform this prospecting. An important issue in designing such systems is the sampling strategy: how do the robots choose where to prospect next? This paper argues that a strategy called Adaptive Cluster Sampling (ACS) has a number of desirable properties: compared to conventional strategies, (1) it reduces the total mission time and energy consumption of a team of robots, and (2) returns a higher mineral yield and more information about the prospected region by directing exploration towards areas of high mineral density, thus providing detailed maps of the boundaries of such areas. Due to the adaptive nature of the sampling scheme, it is not immediately obvious how the resulting sampled data can be used to provide an unbiased, lowvariance estimate of the regional mineral density. This paper therefore investigates new mineral density estimators, which have lower error than previouslydeveloped estimators; they are derived from the older estimators via a process called RaoBlackwellization. Since the efficiency of estimators depends on the type of mineralogical population sampled, the population characteristics that favor ACS estimators are also analyzed. The ACS scheme and our new estimators are evaluated empirically in a detailed simulation of the prospecting task, and the quantitative results show that our approach can yield more minerals with less resources and provide more accurate mineral density estimates than previous methods.
 Integrated Robot Planning and Control with Extended Kohonen Maps.
Kian Hsiang Low.
M.Sc. Thesis, Department of Computer Science, School of Computing, National University of Singapore, Jul 2002.
Singapore Computer Society Prize for best M.Sc. Thesis 20022003
Abstract. The problem of goaldirected, collisionfree motion in a complex, unpredictable environment can be solved by tightly integrating highlevel deliberative planning with lowlevel reactive control. This thesis presents two such architectures for a nonholonomic mobile robot. To achieve realtime performance, reactive control capabilities have to be fully realized so that the deliberative planner can be simplified. These architectures are enriched with reactive target reaching and obstacle avoidance modules. Their target reaching modules use indirectmapping Extended Kohonen Map to provide finer and smoother motion control than directmapping methods. While one architecture fuses these modules indirectly via command fusion, the other one couples them directly using cooperative Extended Kohonen Maps, enabling the robot to negotiate unforeseen concave obstacles. The planner for both architectures use a slippery cells technique to decompose the free workspace into fewer cells, thus reducing search time. Any two points in the cell can still be traversed by reactive motion.
 Mobile Robots That Learn to Navigate.
Kian Hsiang Low.
Honors Thesis, Department of Computer Science,
School of Computing, National University of Singapore, Apr 2001.
Abstract. A sensorimotor controller has been implemented to enable a mobile robot to learn its motion control autonomously and perform simple targetreaching movements. This controller is able to perform fine motion by reducing its selfpositioning error and also, reach a designated target location with minimum delay. The control architecture is in the form of a neural network known as the SelfOrganizing Map. Besides implementing the motor control and the online learning algorithms, the essentiality of a prelearning phase is also evaluated. Then, we explore the possibility of incorporating a novel concept known as Local Linear Smoothing into our batch training algorithm; this notion allows the elimination of the boundary bias phenomenon. Lastly, we suggest a simple approach to learning in an obstacleridden environment.
This document, research.html, has been accessed 53406 times since 26Jun21 15:35:21 SGT.
This is the 34th time it has been accessed today.
A total of 17004 different hosts have accessed this document in the
last 1028 days; your host, ec24420197224.compute1.amazonaws.com, has accessed it 1 times.
If you're interested, complete statistics for
this document are also available, including breakdowns by toplevel
domain, host name, and date.

TRUSTED MODEL/DATA SHARING AND DATA VALUATION
PROJECT DURATION : Oct 2018  Present
PROJECT AFFILIATION

Trusted Collaborative Machine Learning (Trusted CollabML) Lab, NUS Institute of Data Science (IDS) (Collaborator: SeeKiong Ng)

NUS Centre for Research in Privacy Technologies (NCRiPT) Centre (Collaborator: Mohan Kankanhalli)
MEDIA NEWS :

NCRiPT Public Seminar (29 Dec 2020)  Trusted Data Sharing: Incentivizing Collaboration and Rights to be Forgotten (aka Unlearning) in Machine Learning

NCRiPT Blog (3 Jan 2020)  'Sharing, without oversharing, in collective machine learning'
PROJECT FUNDING
 AI Singapore Research Programme : Toward Trustable Modelcentric Sharing for Collaborative Machine Learning, S$8,401,002.40, Apr 2021  Mar 2025
PROBLEM MOTIVATION
 Incentivizing Collaboration for Trusted Data Sharing. Collaborative machine learning (ML) is an appealing paradigm to build highquality ML models. While an individual party may have limited data, it is possible to build improved, highquality ML models by training on the aggregated data from many parties. For example, in healthcare, a hospital or healthcare firm whose data diversity and quantity are limited due to its small patient base can draw on data from other hospitals and firms to improve the prediction of some disease progression (e.g., diabetes). This collaboration can be encouraged by a government agency, such as the National Institute of Health in the United States.
In precision agriculture, a farmer with limited land area and sensors can combine his collected data with the other farmers to improve the modeling of the effect of various influences (e.g., weather, pest) on his crop yield. Such data sharing also benefits other application domains, including real estate in which a property agency can pool together its limited transactional data with that of the other agencies to improve the prediction of property prices.
However, any party would have incurred some nontrivial cost to collect its data. So, they would not altruistically donate their data and risk losing their competitive edge.
These parties will be motivated to share their data when given enough incentives,
such as a guaranteed benefit from the collaboration and a fair higher reward from contributing more valuable data.
 Collective Learning/Fusion of Heterogeneous, BlackBox ML Models. Practical scenarios that involve learning in complex environments often require the collaboration of multiple experts operating concurrently on different subdomains. Motivated under this context, collective learning is an emerging study of a distributed paradigm where each local expert learns independently from its data and exchange knowledge with others to achieve better performance.
Existing collective learning literature, however, usually assumes perfect clarity of local expert models, which entails fully transparent model architectures and publicly accessible local data used to train these experts. To facilitate model fusion, local experts are further expected to have employed a homogeneous design with limited freedom in their choices of parameters. Despite enabling collective learning, these restrictive assumptions have imposed a rigidity on the algorithmic level that is generally undesirable for practical purposes. For example, applications learning from private medical records are often prohibited from publicizing sensitive patient information; model architectures in confidential domains such as financial forecasting are preferably kept undisclosed to guard against adversarial attacks. As such, it is unrealistic for a collaboration scheme among these experts to presume prior understanding of their behaviors, much less subjecting them to conceptual homogeneity.
Another central issue of collective learning is the computational and communication bottleneck arising from having one single or a few central servers to coordinate the collective agents. In practice, such a centralized collective architecture also often leads to having undesirable choke points of failure in the system.
 MultiTask Model Fusion of BlackBox Experts.
In various disciplines such as environmental sensing, traffic monitoring, and healthcare analytics, observational data that describe the same phenomenon or concept are often acquired from multiple experiments, which are conducted on different subjects. The data that they generated would therefore have different distributions or statistical properties. In practice, due to privacy concerns, data collected from different acquisition frameworks (e.g., sensors and/or experiments) are also private and cannot be shared among themselves. This creates private datasets of the same phenomenon, where each is used to train a separate model from a local perspective. For example, in clinical research, patient information is often recorded across different institutions, which do not share data with each other due to protect the patient's sensitive information. As such, each institution can only model the patient population using data collected from a single demographic region, which might not generalize well to others.
Furthermore, in settings with strict security requirements, parameters of a local model also need to be kept private due to a recently discovered threat of adversarial ML attack, which essentially makes it a black box to others. This violates the model transparency requirement of existing distributed modeling works addressing this data federation issue, which results in the blackbox challenge.
In addition, the existing works in federated learning assumes that local models were trained to solve the same task (albeit with different data). Their proposed methods in fact do not isolate taskspecific (irrelevant knowledge) from taskagnostic information (relevant knowledge) which are implicitly entangled in the parameter representation of each model. This entanglement presents a problem in multitask setting since it does not tell us which part of an existing model is relevant to a new task and which part is not. This is also an issue in a remotely relevant literature of meta learning which tackles the model adaptation problem in multitask scenarios from a different technical setting that does not accommodate blackbox and pretrained models with private training data.
PROPOSED METHODOLOGY
 Incentivizing Collaboration for Trusted Data Sharing. We propose to value each party's contributed data and design an incentiveaware reward scheme to give each party a separate ML model as a reward (in short, model reward) accordingly. We use only model rewards and exclude monetary compensations as (a) in the abovementioned applications, every party such as a hospital is mainly interested in improving model quality for unlimited future test predictions; (b) there may not be a feasible and clear source of monetary revenue to compensate participants (e.g., due to restrictions, the government may not be able to pay firms using tax revenues); and (c) if parties have to pay to participate in the collaboration, then the total payment is debatable and many may lack funding while still having valuable data.
 How then should we value a party's data and its effect on model quality?
To answer this first question, we propose a valuation method based on the informativeness of data. In particular, more informative data induces a greater reduction in the uncertainty of the model parameters, hence improving model quality.
In contrast, existing data valuation methods measure model quality via its validation accuracy, which calls for a tedious or even impossible process of needing all parties to agree on a common validation dataset. Inaccurate valuation can also arise due to how a party's test queries, which are likely to change over time, differ from the validation dataset. Our data valuation method does not make any assumption about the current or future distribution of test queries.
 Next, how should we design a reward scheme to decide the values of model rewards for incentivizing a collaboration?
Intuitively, a party will be motivated to collaborate if it can receive a better ML model than others who have contributed less valuable data, and than what it can build alone or get from collaborating separately with some parties.
Also, the parties often like to maximize the total benefit from the collaboration.
These incentives appear related to solution concepts (fairness, individual rationality, stability, and group welfare) from cooperative game theory (CGT), respectively. However, as CGT assumptions are restrictive for the uniquely freely replicable nature of our model reward, these CGT concepts have to be adapted for defining incentives for our model reward. We then design a novel model reward scheme to provide these incentives.
 Finally, how can we realize the values of the model rewards decided by our scheme?
How should we modify the model rewards or the data used to train them?
An obvious approach to control the values of the model rewards is to select and only train on subsets of the aggregated data. However, this requires considering an exponential number of discrete subsets of data, which is intractable for large datasets such as medical records. We avoid this issue by injecting noise into the aggregated data from multiple parties instead. The value of each party's model reward can then be realized by simply optimizing the continuous noise variance parameter.
Current efforts on designing incentive mechanisms for federated learning can build upon our work presented here. For future work, we plan to address privacy preservation in our reward scheme. The noise injection method used for realizing the model rewards in this work is closely related to the Gaussian mechanism of differential privacy. This motivates us to explore how the injected noise will affect privacy in the model rewards.
 Collective Learning/Fusion of Heterogeneous, BlackBox ML Models. We propose a novel collective learning platform for blackbox fusion that addresses the following challenges: (a) performing fusion without access to the blackbox training data and architectures, (b) performing fusion when the blackbox models are not permanently available, and (c) avoiding centralized bottlenecks and risk of failure for largescale fusion with numerous blackbox experts.
To achieve this, we first develop a collective fusion paradigm that allows blackbox experts to interact and learn the predictive behaviors of one another, which are then succinctly encoded into information summaries with constant memory footprint for efficient communication and assimilation. A decentralized communication algorithm is further developed to regulate the propagation flow of these local summaries to optimize the expected improvement rate of the experts while guaranteeing that they reach a consensus upon convergence.
 MultiTask Model Fusion of BlackBox Experts. We propose a fusion paradigm that represents each blackbox expert as a taskdependent distribution over an infinite spectrum of taskagnostic predictive prototypes. The prototypes intuitively represent the taskagnostic information, which can be transferred among tasks to corroborate their statistical strength, whereas their distribution encodes domainspecific information that needs to be adapted to suit a new task.
PUBLICATIONS
 Collaborative Bayesian Optimization with Fair Regret.
Rachael Hwee Ling Sim, Yehong Zhang, Kian Hsiang Low & Patrick Jaillet.
In Proceedings of the 38th International Conference on Machine Learning (ICML21), pages 96919701, Jul 1824, 2021.
21.5% acceptance rate
Abstract. Bayesian optimization (BO) is a popular tool for optimizing complex and costlytoevaluate blackbox objective functions. To further reduce the number of function evaluations, any party performing BO may be interested to collaborate with others to optimize the same objective function concurrently. To do this, existing BO algorithms have considered optimizing a batch of input queries in parallel and provided theoretical bounds on their cumulative regret reflecting inefficiency. However, when the objective function values are correlated with realworld rewards (e.g., money), parties may be hesitant to collaborate if they risk incurring larger cumulative regret (i.e., smaller realworld reward) than others. This paper shows that fairness and efficiency are both necessary for the collaborative BO setting. Inspired by social welfare concepts from economics, we propose a new notion of regret capturing these properties and a collaborative BO algorithm whose convergence rate can be theoretically guaranteed by bounding the new regret, both of which share an adjustable parameter for trading off between fairness vs. efficiency. We empirically demonstrate the benefits (e.g., increased fairness) of our algorithm using synthetic and realworld datasets.
 Model Fusion for Personalized Learning.
Chi Thanh Lam, Trong Nghia Hoang, Kian Hsiang Low & Patrick Jaillet.
In Proceedings of the 38th International Conference on Machine Learning (ICML21), pages 59485958, Jul 1824, 2021.
21.5% acceptance rate
Abstract. Production systems operating on a growing domain of analytic services often require generating warmstart solution models for emerging tasks with limited data. One potential approach to address this challenge is to adopt meta learning to generate a base model that can be adapted to solve unseen tasks with minimal finetuning. This however requires the training processes of previous solution models of existing tasks to be synchronized. This is not possible if these models were pretrained separately on private data owned by different entities and cannot be synchronously retrained. To accommodate for such scenarios, we develop a new personalized learning framework that synthesizes customized models for unseen tasks via fusion of independently pretrained models of related tasks. We establish performance guarantee for the proposed framework and demonstrate its effectiveness on both synthetic and real datasets.
 AID: Active Distillation Machine to Leverage PreTrained BlackBox Models in Private Data Settings.
Trong Nghia Hoang, Shenda Hong, Cao Xiao, Kian Hsiang Low & Jimeng Sun.
In Proceedings of the 30th The Web Conference (WWW'21), pages 3569–3581, Apr 1923, 2021.
20.6% acceptance rate
Abstract. This paper presents an active distillation method for a local institution (e.g., hospital) to find the best queries within its given budget to distill an onserver blackbox model's predictive knowledge into a local surrogate with transparent parameterization. This allows local institutions to understand better the predictive reasoning of the blackbox model in its own local context or to further customize the distilled knowledge with its private dataset that cannot be centralized and fed into the server model. The proposed method thus addresses several challenges of deploying machine learning in many industrial settings (e.g., healthcare analytics) with strong proprietary constraints. These include: (1) the opaqueness of the server model’s architecture which prevents local users from understanding its predictive reasoning in their local data contexts; (2) the increasing cost and risk of uploading local data on the cloud for analysis; and (3) the need to customize the server model with private onsite data. We evaluated the proposed method on both benchmark and realworld healthcare data where significant improvements over existing local distillation methods were observed. A theoretical analysis of the proposed method is also presented.
 Federated Bayesian Optimization via Thompson Sampling.
Zhongxiang Dai, Kian Hsiang Low & Patrick Jaillet.
In Advances in Neural Information Processing Systems 33: 34th Annual Conference on Neural Information Processing Systems (NeurIPS'20), pages 96879699, Dec 612, 2020.
20.1% acceptance rate
Abstract. Bayesian optimization (BO) is a prominent method for optimizing expensivetocompute blackbox functions. The massive computational capability of edge devices such as mobile phones, coupled with privacy concerns, has led to immense recent interest in federated learning (FL), which focuses on collaborative training of deep neural networks (DNN) via firstorder optimization techniques. However, some common machine learning tasks such as hyperparameter tuning of DNN lack access to gradients and thus require zerothorder optimization (blackbox optimization). This hints at the considerable potential of extending BO to the FL setting (FBO), to allow agents to collaborate in these blackbox optimization tasks. Here, we introduce federated Thompson sampling (FTS), which overcomes a number of key challenges of FBO and FL in a principled way: We (a) use random Fourier features to approximate the Gaussian process surrogate model used in BO which naturally produces the parameters to be exchanged between agents, (b) design FTS based on Thompson sampling which significantly reduces the number of parameters to be exchanged, and (c) provide a theoretical convergence guarantee that is robust against heterogeneous agents which is a major challenge in FL and FBO. We empirically demonstrate the effectiveness of FTS in terms of communication efficiency, computational efficiency and practical performance.
 Variational Bayesian Unlearning.
Quoc Phong Nguyen, Kian Hsiang Low & Patrick Jaillet.
In Advances in Neural Information Processing Systems 33: 34th Annual Conference on Neural Information Processing Systems (NeurIPS'20), pages 1602516036, Dec 612, 2020.
20.1% acceptance rate
Abstract. This paper studies the problem of approximately unlearning a Bayesian model from a small subset of the training data to be erased. We frame this problem as one of minimizing the KullbackLeibler distance between the approximate posterior belief of model parameters after directly unlearning from the erased data vs. the exact posterior belief from retraining with remaining data. Using the variational inference (VI) framework, we show that it is equivalent to minimizing an evidence upper bound which trades off between fully unlearning from erased data vs. not entirely forgetting the posterior belief given the full data (i.e., including the remaining data); the latter prevents catastrophic unlearning that can render the model useless. In model training with VI, only an approximate (instead of exact) posterior belief given the full data can be obtained, which makes unlearning even more challenging. We propose two novel tricks to tackle this challenge. We empirically demonstrate our unlearning methods on Bayesian models such as sparse Gaussian process and logistic regression using synthetic and realworld datasets.
 Collaborative Machine Learning with IncentiveAware Model Rewards.
Rachael Hwee Ling Sim, Yehong Zhang, Mun Choon Chan & Kian Hsiang Low.
In Proceedings of the 37th International Conference on Machine Learning (ICML20), pages 89278936, Jun 1218, 2020.
21.8% acceptance rate
Abstract. Collaborative machine learning (ML) is an appealing paradigm to build highquality ML models by training on the aggregated data from many parties. However, these parties are only willing to share their data when given enough incentives, such as a guaranteed fair reward based on their contributions. This motivates the need for measuring a party's contribution and designing an incentiveaware reward scheme accordingly. This paper proposes to value a party's reward based on Shapley value and information gain on model parameters given its data. Subsequently, we give each party a model as a reward. To formally incentivize the collaboration, we define some desirable properties (e.g., fairness and stability) which are inspired by cooperative game theory but adapted for our model reward that is uniquely freely replicable. Then, we propose a novel model reward scheme to satisfy fairness and trade off between the desirable properties via an adjustable parameter. The value of each party's model reward determined by our scheme is attained by injecting Gaussian noise to the aggregated training data with an optimized noise variance. We empirically demonstrate interesting properties of our scheme and evaluate its performance using synthetic and realworld datasets.
 Learning TaskAgnostic Embedding of Multiple BlackBox Experts for MultiTask Model Fusion.
Trong Nghia Hoang, Chi Thanh Lam, Kian Hsiang Low & Patrick Jaillet.
In Proceedings of the 37th International Conference on Machine Learning (ICML20), pages 42824292, Jun 1218, 2020.
21.8% acceptance rate
Abstract. Model fusion is an emerging study in collective learning where heterogeneous experts with private data and learning architectures need to combine their blackbox knowledge for better performance. Existing literature achieves this via a local knowledge distillation scheme that transfuses the predictive patterns of each pretrained expert onto a whitebox imitator model, which can be incorporated efficiently into a global model. This scheme however does not extend to multitask scenarios where different experts were trained to solve different tasks and only part of their distilled knowledge is relevant to a new task. To address this multitask challenge, we develop a new fusion paradigm that represents each expert as a distribution over a spectrum of predictive prototypes, which are isolated from taskspecific information encoded within the prototype distribution. The taskagnostic prototypes can then be reintegrated to generate a new model that solves a new task encoded with a different prototype distribution. The fusion and adaptation performance of the proposed framework is demonstrated empirically on several realworld benchmark datasets.
 Collective Model Fusion for Multiple BlackBox Experts.
Quang Minh Hoang, Trong Nghia Hoang, Kian Hsiang Low & Carleton Kingsford.
In Proceedings of the 36th International Conference on Machine Learning (ICML19), pages 27422750, Long Beach, CA, Jun 915, 2019.
22.6% acceptance rate
Abstract. Model fusion is a fundamental problem in collective machine learning (ML) where independent experts with heterogeneous learning architectures are required to combine expertise to improve predictive performance. This is particularly challenging in informationsensitive domains (e.g., medical records in healthcare analytics) where experts do not have access to each other's internal architecture and local data. To address this challenge, this paper presents the first collective model fusion framework for multiple experts with heterogeneous blackbox architectures. The proposed method will enable this by addressing the following key issues of how blackbox experts interact to understand the predictive behaviors of one another; how these understandings can be represented and shared efficiently among themselves; and how the shared understandings can be combined to generate highquality consensus prediction. The performance of the resulting framework is analyzed theoretically and demonstrated empirically on several datasets.
PRESENTATIONS
 Trusted Data Sharing: Incentivizing Collaboration and Rights to be Forgotten (aka Unlearning) in Machine Learning.
Kian Hsiang Low.
Invited keynote speaker at the 20th IEEE/WIC/ACM International Conference on Web Intelligence and Intelligent Agent Technology, Melbourne, Australia, Dec 1417, 2021.
 Collective Online Learning and Model Fusion in Large Multiagent Systems.
Kian Hsiang Low.
Invited keynote speaker at the 2nd International Symposium on MultiRobot and MultiAgent Systems, Rutgers University, New Brunswick, NJ, USA, Aug 2223, 2019.
AUTOMATED MACHINE LEARNING : BAYESIAN OPTIMIZATION
PROJECT DURATION : Feb 2016  Present
PROJECT AFFILIATION

Temasek Life Sciences Laboratory (TLL) (Collaborator: Nam Hai Chua)

SingaporeMIT Alliance for Research and Technology (SMART) Future Urban Mobility (FM) IRG (Collaborator: Patrick Jaillet, MIT)

SensorEnhanced Social Media (SeSaMe) Centre (Collaborator: Mohan Kankanhalli)
MEDIA NEWS

AI Singapore Spotlight SG AI Researchers (1 Sep 2020)  'Discovering the Science behind Hyperparameter Tuning'

8th QuantumBlack Singapore Meetup (8 July 2020)  'Probabilistic Machine Learning'

NUS School of Computing Faculty Feature (28 May 2019)  'What Bayesian Optimization can teach us about baking better cookies and more'
PROJECT FUNDING
 RIE2020 AME Programmatic Fund : Learning with Less Data, SGD $1,218,600, Apr 2021  Mar 2024
 DSTA Project Agreement : Tactics Discovery and Recommendation, SGD $1,143,120, Jul 2021  Aug 2023
 MOE AcRF Tier 1 Reimagine Research Scheme Funding : Scalable AI Phenome Platform towards FastForward Plant Breeding (Machine Learning),
SGD $348,600, Mar 2021  Mar 2024
 RIE2020 AME IAFPP : High Performance Precision Agriculture (HiPPA) System, S$1,197,960, Mar 2020  Feb 2024
 MOE AcRF Tier 2 Grant : Scaling up Gaussian Process Predictive Models for Big Data, SGD $737,461, Jul 2017  Jul 2020
 SMART Subaward Agreement  FM IRG :
Automatic Probabilistic Machine Learning for Traffic Modeling and Prediction,
SGD $184,999.20, Apr 2017  May 2021
 Research Collaboration Agreement with Panasonic R&D Center Singapore : Hyperparameters Tuning using Bayesian Optimization, SGD $69,336, Mar 2016  Mar 2017
PROBLEM MOTIVATION
 Batch Bayesian Optimization. Bayesian optimization (BO) has recently gained considerable traction due to its capability of finding the global maximum of a highly complex (e.g., nonconvex, no closedform expression nor derivative), noisy blackbox objective function with a limited budget of (often costly) function evaluations, consequently witnessing its use in an increasing diversity of application domains such as robotics, environmental sensing/monitoring, automated machine learning, among others.
A number of acquisition functions (e.g., probability of improvement or expected improvement over the currently found maximum, entropybased, and upper confidence bound (UCB)) have been devised to perform BO: They repeatedly select an input for evaluating/querying the blackbox function (i.e., until the budget is depleted) that intuitively trades off between sampling where the maximum is likely to be given the current, possibly imprecise belief of the function modeled by a Gaussian process (GP) (i.e., exploitation) vs. improving the GP belief of the function over the entire input domain (i.e., exploration) to guarantee finding the global maximum.
The rapidly growing affordability and availability of hardware resources (e.g., computer clusters, sensor networks, robot teams/swarms) have motivated the recent development of BO algorithms that can repeatedly select a batch of inputs for querying the blackbox function in parallel instead. Such batch/parallel BO algorithms can be classified into two types: On one extreme, batch BO algorithms like multipoints expected improvement, parallel predictive entropy search, and the parallel knowledge gradient method jointly optimize the batch of inputs and hence scale poorly in the batch size.
On the other extreme, greedy batch BO algorithms boost the scalability by selecting the inputs of the batch one at a time. We argue that such a highly suboptimal approach to gain scalability is an overkill: In practice, each function evaluation is often much more computationally and/or economically costly (e.g., hyperparameter tuning for deep learning, drug testing on human subjects), which justifies dedicating more time to obtain better BO performance.
 Nonmyopic Bayesian Optimization. The fundamental challenge of integrated planning and learning is to design an autonomous agent that can plan its actions to maximize its expected total rewards while interacting with an unknown task environment. Recent research efforts tackling this challenge have progressed from the use of simple Markov models assuming discretevalued, independent observations to that of a rich class of Bayesian nonparametric Gaussian process (GP) models characterizing continuousvalued, correlated observations in order to represent the latent structure of complex, possibly noisy task environments with higher fidelity. Such a challenge is posed by the problem of Bayesian optimization (BO).
Its objective is to select and gather the most informative (possibly noisy) observations for finding the global maximum of an unknown, highly complex (e.g., nonconvex, no closedform expression nor derivative) objective function (i.e., task environment) modeled by a GP given a sampling budget (e.g., number of costly function evaluations). The rewards of a BO agent are defined using an improvementbased (e.g., probability of improvement or expected improvement over currently found maximum), entropybased, or upper confidence bound (UCB) acquisition function. A limitation of most BO algorithms is that they are myopic. To overcome this limitation, approximation algorithms for nonmyopic adaptive BO have been proposed, but their performances are not theoretically guaranteed.
PROPOSED METHODOLOGY
 Batch Bayesian Optimization. To tackle the first problem, we show that it is in fact possible to jointly optimize the batch of inputs and still preserve scalability in the batch size by giving practitioners the flexibility to trade off BO performance for time efficiency.
To achieve this, we first observe that, interestingly, batch BO can be perceived as a cooperative multiagent decision making problem whereby each agent optimizes a separate input of the batch while coordinating with the other agents doing likewise.
To the best of our knowledge, this has not been considered in the BO literature.
In particular, if batch BO can be framed as some known class of multiagent decision making problems, then it can be solved efficiently and scalably by the latter's stateoftheart solvers.
The key technical challenge would therefore be to investigate how batch BO can be cast as one of such to exploit its advantage of scalability in the number of agents (hence, batch size) while at the same time theoretically guaranteeing the resulting BO performance.
To tackle the above challenge, our first work presents a novel distributed batch BO algorithm that, in contrast to greedy batch BO algorithms,
can jointly optimize a batch of inputs and, unlike the batch BO algorithms, still preserve scalability in the batch size.
To realize this, we generalize GPUCB to a new batch variant amenable to a Markov approximation, which can then be naturally formulated as a multiagent distributed constraint optimization problem (DCOP) in order to fully exploit the efficiency of its stateoftheart solvers for achieving linear time in the batch size.
Our proposed distributed batch GPUCB (DBGPUCB) algorithm offers practitioners the flexibility to trade off between the approximation quality and time efficiency by varying the Markov order. We provide a theoretical guarantee for the convergence rate of our DBGPUCB algorithm via bounds on its cumulative regret. We empirically evaluate the cumulative regret incurred by our DBGPUCB algorithm and its scalability in the batch size on synthetic benchmark objective functions and a realworld optimization problem.
 Nonmyopic Bayesian Optimization. To address the second problem, our second work presents a novel nonmyopic adaptive Gaussian process planning (GPP) framework endowed with a general class of Lipschitz continuous reward functions that can unify some active learning and BO criteria (e.g., UCB) and offer practitioners some flexibility to specify their desired choices for defining new tasks/problems. In particular, it utilizes a principled Bayesian sequential decision problem framework for jointly and naturally optimizing the explorationexploitation tradeoff, consequently allowing planning and learning to be integrated seamlessly and performed simultaneously instead of separately. In general, the resulting induced GPP policy cannot be derived exactly due to an uncountable set of candidate observations. A key contribution of our work here thus lies in exploiting the Lipschitz continuity of the reward functions to solve for a nonmyopic adaptive εoptimal GPP (εGPP) policy given an arbitrarily userspecified loss bound ε. To plan in real time, we further propose an asymptotically optimal, branchandbound anytime variant of εGPP with performance guarantee. Finally, we empirically evaluate the performances of our εGPP policy and its anytime variant in BO and an energy harvesting task on simulated and realworld environmental fields.
PUBLICATIONS
 TrustedMaximizers Entropy Search for Efﬁcient Bayesian Optimization.
Quoc Phong Nguyen, Zhaoxuan Wu, Kian Hsiang Low & Patrick Jaillet^{}.
In Proceedings of the 37th Conference on Uncertainty in Artificial Intelligence (UAI21), pages 14861495, Jul 2730, 2021.
26.5% acceptance rate
Abstract. Informationbased Bayesian optimization (BO) algorithms have achieved stateoftheart performance in optimizing a blackbox objective function. However, they usually require several approximations or simplifying assumptions (without clearly understanding their effects on the BO performance) and/or their generalization to batch BO is computationally unwieldy, especially with an increasing batch size. To alleviate these issues, this paper presents a novel trustedmaximizers entropy search (TES) acquisition function: It measures how much an input query contributes to the information gain on the maximizer over a ﬁnite set of trusted maximizers, i.e., inputs optimizing functions that are sampled from the Gaussian process posterior belief of the objective function. Evaluating TES requires either only a stochastic approximation with sampling or a deterministic approximation with expectation propagation, both of which are investigated and empirically evaluated using synthetic benchmark objective functions and realworld optimization problems, e.g., hyperparameter tuning of a convolutional neural network and synthesizing ‘physically realizable’ faces to fool a blackbox face recognition system. Though TES can naturally be generalized to a batch variant with either approximation, the latter is amenable to be scaled to a much larger batch size in our experiments.
 ValueatRisk Optimization with Gaussian Processes.
Quoc Phong Nguyen^{}, Zhongxiang Dai^{}, Kian Hsiang Low & Patrick Jaillet^{}.
In Proceedings of the 38th International Conference on Machine Learning (ICML21), pages 80638072, Jul 1824, 2021.
21.5% acceptance rate
Abstract. Valueatrisk (VaR) is an established measure to assess risks in critical realworld applications with random environmental factors. This paper presents a novel VaR upper confidence bound (VUCB) algorithm for maximizing the VaR of a blackbox objective function with the first noregret guarantee. To realize this, we first derive a confidence bound of VaR and then prove the existence of values of the environmental random variable (to be selected to achieve no regret) such that the confidence bound of VaR lies within that of the objective function evaluated at such values. Our VUCB algorithm empirically demonstrates stateoftheart performance in optimizing synthetic benchmark functions, a portfolio optimization problem, and a simulated robot task.
 Collaborative Bayesian Optimization with Fair Regret.
Rachael Hwee Ling Sim^{}, Yehong Zhang^{}, Kian Hsiang Low & Patrick Jaillet^{}.
In Proceedings of the 38th International Conference on Machine Learning (ICML21), pages 96919701, Jul 1824, 2021.
21.5% acceptance rate
Abstract. Bayesian optimization (BO) is a popular tool for optimizing complex and costlytoevaluate blackbox objective functions. To further reduce the number of function evaluations, any party performing BO may be interested to collaborate with others to optimize the same objective function concurrently. To do this, existing BO algorithms have considered optimizing a batch of input queries in parallel and provided theoretical bounds on their cumulative regret reflecting inefficiency. However, when the objective function values are correlated with realworld rewards (e.g., money), parties may be hesitant to collaborate if they risk incurring larger cumulative regret (i.e., smaller realworld reward) than others. This paper shows that fairness and efficiency are both necessary for the collaborative BO setting. Inspired by social welfare concepts from economics, we propose a new notion of regret capturing these properties and a collaborative BO algorithm whose convergence rate can be theoretically guaranteed by bounding the new regret, both of which share an adjustable parameter for trading off between fairness vs. efficiency. We empirically demonstrate the benefits (e.g., increased fairness) of our algorithm using synthetic and realworld datasets.
 Topk Ranking Bayesian Optimization.
Quoc Phong Nguyen, Sebastian Tay, Kian Hsiang Low & Patrick Jaillet.
In Proceedings of the 35th AAAI Conference on Artificial Intelligence (AAAI21), pages 91359143, Feb 29, 2021.
21.4% acceptance rate
Abstract. This paper presents a novel approach to topk ranking Bayesian optimization (topk ranking BO) which is a practical and significant generalization of preferential BO to handle topk ranking and tie/indifference observations. We first design a surrogate model that is not only capable of catering to the above observations, but is also supported by a classic random utility model. Another equally important contribution is the introduction of the first informationtheoretic acquisition function in BO with preferential observation called multinomial predictive entropy search (MPES) which is flexible in handling these observations and optimized for all inputs of a query jointly. MPES possesses superior performance compared with existing acquisition functions that select the inputs of a query one at a time greedily. We empirically evaluate the performance of MPES using several synthetic benchmark functions, CIFAR10 dataset, and SUSHI preference dataset.
 An InformationTheoretic Framework for Unifying Active Learning Problems.
Quoc Phong Nguyen, Kian Hsiang Low & Patrick Jaillet.
In Proceedings of the 35th AAAI Conference on Artificial Intelligence (AAAI21), pages 91269134, Feb 29, 2021.
21.4% acceptance rate
Abstract. This paper presents an informationtheoretic framework for unifying active learning problems: level set estimation (LSE), Bayesian optimization (BO), and their generalized variant. We first introduce a novel active learning criterion that subsumes an existing LSE algorithm and achieves stateoftheart performance in LSE problems with a continuous input domain. Then, by exploiting the relationship between LSE and BO, we design a competitive informationtheoretic acquisition function for BO that has interesting connections to upper confidence bound and maxvalue entropy search (MES). The latter connection reveals a drawback of MES which has important implications on not only MES but also on other MESbased acquisition functions. Finally, our unifying informationtheoretic framework can be applied to solve a generalized problem of LSE and BO involving multiple level sets in a dataefficient manner. We empirically evaluate the performance of our proposed algorithms using synthetic benchmark functions, a realworld dataset, and in hyperparameter tuning of machine learning models.
 Efficient Exploration of Reward Functions in Inverse Reinforcement Learning via Bayesian Optimization.
Sreejith Balakrishnan, Quoc Phong Nguyen, Kian Hsiang Low & Harold Soh.
In Advances in Neural Information Processing Systems 33: 34th Annual Conference on Neural Information Processing Systems (NeurIPS'20), pages 41874198, Dec 612, 2020.
20.1% acceptance rate
Abstract. In this paper, we focus on the problem of Inverse Reinforcement Learning (IRL), which is relevant for a variety of tasks including value alignment and robot learning from demonstration. Despite significant algorithmic contributions in recent years, IRL remains an illposed problem at its core; multiple reward functions coincide with the observed behavior, and the actual reward function is not identifiable without prior knowledge or supplementary information. Here, we propose Bayesian OptimizationIRL (BOIRL), an IRL framework that identifies multiple solutions that are consistent with the expert demonstrations by efficiently exploring the reward function space. BOIRL achieves this by utilizing Bayesian Optimization along with our newly proposed kernel that (a) projects the parameters of policy invariant reward functions to a single point in a latent space, and (b) ensures that nearby points in the latent space correspond to reward functions that yield similar likelihoods. This projection allows for the use of standard stationary kernels in the latent space to capture the correlations present across the reward function space. Empirical results on synthetic and realworld environments (modelfree and modelbased) show that BOIRL discovers multiple reward functions while minimizing the number of expensive exact policy optimizations.
 Federated Bayesian Optimization via Thompson Sampling.
Zhongxiang Dai, Kian Hsiang Low & Patrick Jaillet.
In Advances in Neural Information Processing Systems 33: 34th Annual Conference on Neural Information Processing Systems (NeurIPS'20), pages 96879699, Dec 612, 2020.
20.1% acceptance rate
Abstract. Bayesian optimization (BO) is a prominent method for optimizing expensivetocompute blackbox functions. The massive computational capability of edge devices such as mobile phones, coupled with privacy concerns, has led to immense recent interest in federated learning (FL), which focuses on collaborative training of deep neural networks (DNN) via firstorder optimization techniques. However, some common machine learning tasks such as hyperparameter tuning of DNN lack access to gradients and thus require zerothorder optimization (blackbox optimization). This hints at the considerable potential of extending BO to the FL setting (FBO), to allow agents to collaborate in these blackbox optimization tasks. Here, we introduce federated Thompson sampling (FTS), which overcomes a number of key challenges of FBO and FL in a principled way: We (a) use random Fourier features to approximate the Gaussian process surrogate model used in BO which naturally produces the parameters to be exchanged between agents, (b) design FTS based on Thompson sampling which significantly reduces the number of parameters to be exchanged, and (c) provide a theoretical convergence guarantee that is robust against heterogeneous agents which is a major challenge in FL and FBO. We empirically demonstrate the effectiveness of FTS in terms of communication efficiency, computational efficiency and practical performance.
 Automated Machine Learning: New Advances on Bayesian Optimization.
Dmitrii Kharkovskii^{}.
Ph.D. Thesis, Department of Computer Science, National University of Singapore, Dec 2020.
Abstract. Recent advances in Bayesian optimization (BO) have delivered a promising suite of tools for optimizing an unknown expensive to evaluate blackbox objective function with a finite budget of evaluations. A significant advantage of BO is its general formulation: BO can be utilized to optimize any blackbox objective function. As a result, BO has been applied in a wide range of applications such as automated machine learning, robotics or environmental monitoring, among others. Furthermore, its general formulation makes BO attractive for deployment in new applications. However, potential new applications can have additional requirements not satisfied by the classical BO setting. In this thesis, we aim to address some of these requirements in order to scale up BO technology for the practical use in new realworld applications.
Firstly, this thesis tackles the problem of data privacy, which is not addressed by the standard setting of BO. Specifically, we consider the outsourced setting where the entity holding the dataset and the entity performing BO are represented by different parties, and the dataset cannot be released nonprivately. For example, a hospital holds a dataset of sensitive medical records and outsources the BO task on this dataset to an industrial AI company. We present the privateoutsourcedGaussian processupper confidence bound (POGPUCB) algorithm, which is the first algorithm for privacypreserving BO in the outsourced setting with a provable performance guarantee. The key idea of our approach is to make the BO performance of our algorithm similar to that of nonprivate GPUCB run using the original dataset, which is achieved by using a random projectionbased transformation that preserves both privacy and the pairwise distances between inputs. Our main theoretical contribution is to show that a regret bound similar to that of the standard GPUCB algorithm can be established for our POGPUCB algorithm. We empirically evaluate the performance of our algorithm with synthetic and realworld datasets.
Secondly, we consider applications of BO for hotspot sampling in spatially varying phenomena. For such applications, we exploit the structure of the spatially varying phenomenon in order to increase the BO lookahead and, as a result, improve the performance of the algorithm and make it more suitable for practical use in realworld scenarios. To do this, we present a principled multistaged Bayesian sequential decision algorithm for nonmyopic adaptive BO that, in particular, exploits macroactions for scaling up to a further lookahead to match up to a larger available budget. To achieve this, we first generalize GPUCB to a new acquisition function defined with respect to a nonmyopic adaptive macroaction policy, which, unfortunately, is intractable to be optimized exactly due to an uncountable set of candidate outputs. The key novel contribution of our work here is to show that it is in fact possible to solve for a nonmyopic adaptive εBayesoptimal macroaction BO (εMacroBO) policy given an arbitrary userspecified loss bound ε via stochastic sampling in each planning stage which requires only a polynomial number of samples in the length of macroactions. To perform nonmyopic adaptive BO in real time, we then propose an asymptotically optimal anytime variant of our εMacroBO algorithm with a performance guarantee. Empirical evaluation on synthetic and realworld datasets shows that our proposed approach outperforms existing stateoftheart algorithms.
Finally, this thesis proposes a blackbox attack for adversarial machine learning based on BO. Since the dimension of the inputs in adversarial learning is usually too high for applying BO directly, our proposed attack applies dimensionality reduction and searches for an adversarial perturbation in a lowdimensional latent space. The key idea of our approach is to automate both the selection of the latent space dimen sion and the search of the adversarial perturbation in the selected latent space by using BO. Additionally, we use Bayesian optimal stopping to boost the query efficiency of our attack. Performance evaluation using image classification datasets shows that our proposed method outperforms the stateoftheart blackbox adversarial attacks.
 R2B2: Recursive ReasoningBased Bayesian Optimization for NoRegret Learning in Games.
Zhongxiang Dai, Yizhou Chen, Kian Hsiang Low, Patrick Jaillet & TeckHua Ho.
In Proceedings of the 37th International Conference on Machine Learning (ICML20), pages 22912301, Jun 1218, 2020.
21.8% acceptance rate
Abstract. This paper presents a recursive reasoning formalism of Bayesian optimization (BO) to model the reasoning process in the interactions between boundedly rational, selfinterested agents with unknown, complex, and costlytoevaluate payoff functions in repeated games, which we call Recursive ReasoningBased BO (R2B2).
Our R2B2 algorithm is general in that it does not constrain the relationship among the payoff functions of different agents and can thus be applied to various types of games such as constantsum, generalsum, and commonpayoff games. We prove that by reasoning at level 2 or more and at one level higher than the other agents, our R2B2 agent can achieve faster asymptotic convergence to no regret than that without utilizing recursive reasoning. We also propose a computationally cheaper variant of R2B2 called R2B2Lite at the expense of a weaker convergence guarantee. The performance and generality of our R2B2 algorithm are empirically demonstrated using synthetic games, adversarial machine learning, and multiagent reinforcement learning.
 Private Outsourced Bayesian Optimization.
Dmitrii Kharkovskii, Zhongxiang Dai & Kian Hsiang Low.
In Proceedings of the 37th International Conference on Machine Learning (ICML20), pages 52315242, Jun 1218, 2020.
21.8% acceptance rate
Abstract. This paper presents the privateoutsourcedGaussian processupper confidence bound (POGPUCB) algorithm, which is the first algorithm for privacypreserving Bayesian optimization (BO) in the outsourced setting with a provable performance guarantee. We consider the outsourced setting where the entity holding the dataset and the entity performing BO are represented by different parties, and the dataset cannot be released nonprivately. For example, a hospital holds a dataset of sensitive medical records and outsources the BO task on this dataset to an industrial AI company.
The key idea of our approach is to make the BO performance of our algorithm similar to that of nonprivate GPUCB run using the original dataset, which is achieved by using a random projectionbased transformation that preserves both privacy and the pairwise distances between inputs. Our main theoretical contribution is to show that a regret bound similar to that of the standard GPUCB algorithm can be established for our POGPUCB algorithm. We empirically evaluate the performance of our POGPUCB algorithm with synthetic and realworld datasets.
 Nonmyopic Gaussian Process Optimization with MacroActions.
Dmitrii Kharkovskii, Chun Kai Ling & Kian Hsiang Low.
In Proceedings of the 23rd International Conference on Artificial Intelligence and Statistics (AISTATS20), pages 45934604, Aug 2628, 2020.
28.7% acceptance rate
Abstract. This paper presents a multistaged approach to nonmyopic adaptive Gaussian process optimization (GPO) for Bayesian optimization (BO) of unknown, highly complex objective functions that, in contrast to existing nonmyopic adaptive BO algorithms, exploits the notion of macroactions for scaling up to a further lookahead to match up to a larger available budget. To achieve this, we generalize GP upper confidence bound to a new acquisition function defined w.r.t. a nonmyopic adaptive macroaction policy, which is intractable to be optimized exactly due to an uncountable set of candidate outputs. The contribution of our work here is thus to derive a nonmyopic adaptive ϵBayesoptimal macroaction GPO (ϵMacroGPO) policy. To perform nonmyopic adaptive BO in real time, we then propose an asymptotically optimal anytime variant of our ϵMacroGPO policy with a performance guarantee. We empirically evaluate the performance of our ϵMacroGPO policy and its anytime variant in BO with synthetic and realworld datasets.
 Bayesian Optimization with Binary Auxiliary Information.
Yehong Zhang, Zhongxiang Dai & Kian Hsiang Low.
In Proceedings of the 35th Conference on Uncertainty in Artificial Intelligence (UAI19), pages 12221232, Tel Aviv, Israel, Jul 2225, 2019.
26.2% acceptance rate (plenary talk)
Subsumes our work on InformationBased MultiFidelity Bayesian Optimization presented in
NeurIPS'17 Workshop on Bayesian Optimization, Long Beach, CA, Dec 9, 2017.
Abstract. This paper presents novel mixedtype Bayesian optimization (BO) algorithms to accelerate the optimization of a target objective function by exploiting correlated auxiliary information of binary type that can be more cheaply obtained, such as in policy search for reinforcement learning and hyperparameter tuning of machine learning models with early stopping. To achieve this, we first propose a mixedtype multioutput Gaussian process (MOGP) to jointly model the continuous target function and binary auxiliary functions. Then, we propose informationbased acquisition functions such as mixedtype entropy search (MTES) and mixedtype predictive ES (MTPES) for mixedtype BO based on the MOGP predictive belief of the target and auxiliary functions. The exact acquisition functions of MTES and MTPES cannot be computed in closed form and need to be approximated. We derive an efficient approximation of MTPES via a novel mixedtype random features approximation of the MOGP model whose crosscorrelation structure between the target and auxiliary functions can be exploited for improving the belief of the global target maximizer using the observations from evaluating these functions. We also propose new practical constraints to relate the global target maximizer to the binary auxiliary functions. We empirically evaluate the performance of MTES and MTPES with synthetic and realworld experiments.
 Bayesian Optimization Meets Bayesian Optimal Stopping.
Zhongxiang Dai, Haibin Yu, Kian Hsiang Low & Patrick Jaillet.
In Proceedings of the 36th International Conference on Machine Learning (ICML19), pages 14961506, Long Beach, CA, Jun 915, 2019.
22.6% acceptance rate
Abstract. Bayesian optimization (BO) is a popular paradigm for optimizing the hyperparameters of machine learning (ML) models due to its sample efficiency. Many ML models require running an iterative training procedure (e.g., stochastic gradient descent). This motivates the question whether information available during the training process (e.g., validation accuracy after each epoch) can be exploited for improving the epoch efficiency of BO algorithms by earlystopping model training under hyperparameter settings that will end up underperforming and hence eliminating unnecessary training epochs. This paper proposes to unify BO (specifically, Gaussian processupper confidence bound (GPUCB)) with Bayesian optimal stopping (BOBOS) to boost the epoch efficiency of BO. To achieve this, while GPUCB is sampleefficient in the number of function evaluations, BOS complements it with epoch efficiency for each function evaluation by providing a principled optimal stopping mechanism for early stopping. BOBOS preserves the (asymptotic) noregret performance of GPUCB using our specified choice of BOS parameters that is amenable to an elegant interpretation in terms of the explorationexploitation tradeoff. We empirically evaluate the performance of BOBOS and demonstrate its generality in hyperparameter optimization of ML models and two other interesting applications.
 Decentralized HighDimensional Bayesian Optimization with Factor Graphs.
Trong Nghia Hoang, Quang Minh Hoang, Ruofei Ouyang & Kian Hsiang Low.
In Proceedings of the 32nd AAAI Conference on Artificial Intelligence (AAAI18), pages 32313238, New Orleans, LA, Feb 28, 2018.
24.55% acceptance rate
Abstract. This paper presents a novel decentralized highdimensional Bayesian optimization (DECHBO) algorithm that, in contrast to existing HBO algorithms, can exploit the interdependent effects of various input components on the output of the unknown objective function f for boosting the BO performance and still preserve scalability in the number of input dimensions without requiring prior knowledge or the existence of a low (effective) dimension of the input space. To realize this, we propose a sparse yet rich factor graph representation of f to be exploited for designing an acquisition function that can be similarly represented by a sparse factor graph and hence be efficiently optimized in a decentralized manner using distributed message passing. Despite richly characterizing the interdependent effects of the input components on the output of f with a factor graph, DECHBO can still guarantee (asymptotic) noregret performance. Empirical evaluation on synthetic and realworld experiments shows that DECHBO outperforms the stateoftheart HBO algorithms.
 DataEfficient Machine Learning with Multiple Output Types and High Input Dimensions.
Yehong Zhang^{}.
Ph.D. Thesis, Department of Computer Science, National University of Singapore, Dec 2017.
Abstract.
Recent research works in machine learning (ML) have focused on learning some target variables of interest to achieve competitive (or stateoftheart) predictive performance in less time but without requiring large quantities of data, which is known as dataefficient ML. This thesis focuses on two important dataefficient ML approaches: active learning (AL) and Bayesian optimization (BO) which, instead of learning passively from a given small set of data, need to select and gather the most informative observations for learning the target variables of interest more accurately given some budget constraints. To advance the stateoftheart of dataefficient ML, novel generalizations of AL and BO algorithms are proposed in this thesis for addressing the issues arising from multiple output types and high input dimensions which are the practical settings in many realworld applications.
In particular, this thesis aims to (a) exploit the auxiliary types of outputs which usually coexist and correlate well with the target output types, and more importantly, are less noisy and/or less tedious to sample for improving the learning performance of the target output type in both AL and BO algorithms and (b) scale up the stateoftheart BO algorithm to high input dimensions. To achieve this, the specific data with multiple output types or high input dimensions is represented using some form of Gaussian process (GP)based probabilistic regression models which allow the predictive uncertainty of the outputs to be formally quantified and consequently exploited for developing efficient AL and BO algorithms.
To achieve above objectives, an AL algorithm of multioutput GP (MOGP) is first developed for minimizing the predictive uncertainty (i.e., posterior joint entropy) of the target output type. In contrast to existing works, our AL problems involve selecting not just the most informative sampling inputs to be observed but also the types of outputs at each selected input for improving the learning performance of only the target output type given a sampling budget. Unfortunately, such an entropy criterion scales poorly in the numbers of candidate sampling inputs and selected observations when optimized. To resolve this issue, we exploit a structure common to sparse MOGP models for deriving a novel AL criterion. Furthermore, we exploit a relaxed form of submodularity property of our new criterion for devising a polynomialtime approximation algorithm that guarantees a constantfactor approximation of that achieved by the optimal set of selected observations. Empirical evaluation on realworld datasets shows that our proposed approach outperforms existing algorithms for AL of MOGP and singleoutput GP models.
Secondly, to boost the BO performance by exploiting the cheaper and/or less noisy observations of some auxiliary functions with varying fidelities, we proposed a novel generalization of predictive entropy search (PES) for multifidelity BO called multifidelity PES (MFPES). In contrast to existing multifidelity BO algorithms, our proposed MFPES algorithm can naturally trade off between exploitation vs. exploration over the target and auxiliary functions with varying fidelities without needing to manually tune any such parameters. To achieve this, we model the unknown target and auxiliary functions jointly as a convolved MOGP (CMOGP) whose convolutional structure is exploited to formally characterize the fidelity of each auxiliary function through its crosscorrelation with the target function. Although the exact acquisition function of MFPES cannot be computed in closed form, we show that it is in fact possible to derive an efficient approximation of MFPES via a novel multioutput random features approximation of the CMOGP model whose crosscorrelation (i.e., multifidelity) structure between the target and auxiliary functions can be exploited for improving the belief of the global target maximizer using the observations from evaluating these functions. Practical constraints are proposed to relate the global target maximizer to that of auxiliary functions. Empirical evaluation on synthetic and realworld experiments shows that MFPES outperforms the stateoftheart multifidelity BO algorithms.
Lastly, to improve the BO performance in realworld applications with high input
dimensions (e.g., computer vision, biology), we generalize PES for highdimensional BO by exploiting an additive structure of the target function. New practical constraints are proposed and approximated efficiently such that the proposed acquisition function of additive PES (addPES) can be optimized independently for each local and lowdimensional input component. The empirical results show that our addPES considerably improves the performance of the stateoftheart highdimensional BO algorithms by using a simple and common setting for optimizing different tested functions with varying input dimensions, which makes it a superior alternative to existing highdimensional BO algorithms.
 Distributed Batch Gaussian Process Optimization.
Erik Daxberger & Kian Hsiang Low.
In Proceedings of the 34th International Conference on Machine Learning (ICML17), pages 951960, Sydney, Australia, Aug 611, 2017.
25.9% acceptance rate
Abstract. This paper presents a novel distributed batch Gaussian process upper confidence bound (DBGPUCB) algorithm for performing batch Bayesian optimization (BO) of highly complex, costlytoevaluate blackbox objective functions. In contrast to existing batch BO algorithms, DBGPUCB can jointly optimize a batch of inputs (as opposed to selecting the inputs of a batch one at a time) while still preserving scalability in the batch size. To realize this, we generalize GPUCB to a new batch variant amenable to a Markov approximation, which can then be naturally formulated as a multiagent distributed constraint optimization problem in order to fully exploit the efficiency of its stateoftheart solvers for achieving linear time in the batch size. Our DBGPUCB algorithm offers practitioners the flexibility to trade off between the approximation quality and time efficiency by varying the Markov order. We provide a theoretical guarantee for the convergence rate of DBGPUCB via bounds on its cumulative regret. Empirical evaluation on synthetic benchmark objective functions and a realworld optimization problem shows that DBGPUCB outperforms the stateoftheart batch BO algorithms.
 Exploiting Decentralized MultiAgent Coordination for LargeScale Machine Learning Problems.
Ruofei Ouyang.
Ph.D. Thesis, Department of Computer Science, National University of Singapore, Dec 2016.
Abstract.
Nowadays, the scale of machine learning problems becomes much larger than before. It raises a huge demand in distributed perception and distributed computation. A multiagent system provides exceptional scalability for problems like active sensing and data fusion. However, many rich characteristics of largescale machine learning problems have not been addressed yet such as large input domain, nonstationarity, and high dimensionality. This thesis identifies the challenges related to these characteristics from multiagent perspective. By exploiting the correlation structure of data in largescale problems, we propose multiagent coordination schemes that can improve the scalability of the machine learning models while preserving the computation accuracy. To elaborate, the machine learning problems we are solving with multiagent coordination techniques are:
 Gaussian process regression. To perform distributed regression on a largescale environmental phenomenon, data compression is often required due to the communication costs. Currently, decentralized data fusion methods encapsulate the data into local summaries based on a fixed support set. However in a largescale field, this fixed support set, acting as a centralized component in the decentralized system, cannot approximate the correlation structure of the entire phenomenon well. It leads to evident losses in data summarization. Consequently, the regression performance will be significantly reduced.
In order to approximate the correlation structure accurately, we propose an agentcentric support set to allow every agent in the data fusion system to choose a possibly different support set and dynamically switch to another one during execution for encapsulating its own data into a local summary which, perhaps surprisingly, can still be assimilated with the other agents’ local summaries into a globally consistent summary. Together with an information sharing mechanism we designed, the new decentralized data fusion methods with agentcentric support set can be applied to regression problems on a much larger environmental phenomenon with high performance.
 Active learning. In the context of environmental sensing, active learning/active sensing is a process of taking observations to minimize the uncertainty in an environmental field. The uncertainty is quantified based on the correlation structure of the phenomenon which is traditionally assumed to be stationary for computational sake. In a largescale environmental field, this stationary assumption is often violated. Therefore, existing active sensing algorithms perform suboptimally for a nonstationary environmental phenomenon.
To the best of our knowledge, our decentralized multirobot active sensing (DECMAS) algorithm is the first work to address nonstationarity issue in the context of active sensing. The uncertainty in the phenomenon is quantified based on the nonstationary correlation structure estimated by Dirichlet process mixture of Gaussian processes. Further, our DECMAS algorithm can efficiently coordinate the exploration of multiple robots to automatically tradeoff between learning the unknown, nonstationary correlation structure and minimizing the uncertainty of the environmental phenomenon. It enables multiagent active sensing techniques to be applied to a largescale nonstationary environmental phenomenon.
 Bayesian optimization. Optimizing an unknown objective function is challenging for traditional optimization methods. Alternatively, in this situation, people use Bayesian optimization which is a modern optimization technique that can optimize a function by only utilizing the observation information (input and output values) collected through simulations. When the input dimension of the function is low, a few simulated observations can generate good result already. However, for high dimensional function, a huge number of observations are required which is impractical when the simulation consumes lots of time and resources.
Fortunately, many high dimensional problems have sparse correlation structure. Our ANOVADCOP work can decompose the correlation structure in the original highdimensional problem into many correlation structures of subsets of dimensions based on ANOVA kernel function. It significantly reduces the size of input space into a collection of lowerdimensional subspaces. Additionally, we reformulate the Bayesian optimization problem as a decentralized constrained optimization problem (DCOP) that can be efficiently solved by multiagent coordination techniques so that it can scale up to problems with hundreds of dimensions.
 Gaussian Process Planning with Lipschitz Continuous Reward Functions: Towards Unifying Bayesian Optimization, Active Learning, and Beyond.
Chun Kai Ling, Kian Hsiang Low & Patrick Jaillet.
In Proceedings of the 30th AAAI Conference on Artificial Intelligence (AAAI16), pages 18601866, Phoenix, AZ, Feb 1217, 2016.
25.75% acceptance rate
Abstract. This paper presents a novel nonmyopic adaptive Gaussian process planning (GPP) framework endowed with a general class of Lipschitz continuous reward functions that can unify some active learning/sensing and Bayesian optimization criteria and offer practitioners some flexibility to specify their desired choices for defining new tasks/problems. In particular, it utilizes a principled Bayesian sequential decision problem framework for jointly and naturally optimizing the explorationexploitation tradeoff. In general, the resulting induced GPP policy cannot be derived exactly due to an uncountable set of candidate observations. A key contribution of our work here thus lies in exploiting the Lipschitz continuity of the reward functions to solve for a nonmyopic adaptive ϵoptimal GPP (ϵGPP) policy. To plan in real time, we further propose an asymptotically optimal, branchandbound anytime variant of ϵGPP with performance guarantee. We empirically demonstrate the effectiveness of our ϵGPP policy and its anytime variant in Bayesian optimization and an energy harvesting task.
PRESENTATIONS
 Learning with Less Data: Automated Machine Learning and Bayesian Optimization.
Kian Hsiang Low.
Invited speaker at the KAUST Research Conference on Robotics and Autonomy, KAUST, Saudi Arabia, Feb 28, 2022.
 Informative Gaussian Process Planning with Lipschitz Continuous Reward Functions: Towards Unifying Adaptive Sampling, Bayesian Optimization, Active Learning, and Beyond.
Kian Hsiang Low.
Invited speaker at the ICRA'18 Workshop on Informative Path Planning and Adaptive Sampling, Brisbane, Australia, May 21, 2018.
Invited speaker at the Symposium on Oceanographic Data Analytics, Norwegian University of Science and Technology, Trondheim, Norway, Nov 2730, 2018.
EXPLORATIONEXPLOITATION DILEMMA IN ACTIVE LEARNING OF GAUSSIAN PROCESSES
PROJECT DURATION : Aug 2013  Present
PROJECT AFFILIATION

SingaporeMIT Alliance for Research and Technology (SMART) Future Urban Mobility (FM) IRG (Collaborator: Patrick Jaillet, MIT)

SensorEnhanced Social Media (SeSaMe) Centre (Collaborator: Mohan Kankanhalli)
PROJECT FUNDING
 RIE2020 AME Programmatic Fund : Learning with Less Data, SGD $1,218,600, Apr 2021  Mar 2024
 Research Collaboration Agreement with InfoCommunications Media Development Authority (IMDA) : Robust and Scalable Computer Vision for Scene Understanding, SGD $260,000, Oct 2020  Oct 2021
 MOE AcRF Tier 2 Grant : Scaling up Gaussian Process Predictive Models for Big Data, SGD $737,461, Jul 2017  Jul 2020
 SMART Subaward Agreements  FM IRG : Spatiotemporal Modeling and Prediction of Traffic Patterns,
SGD $361,456.17, Oct 2011  Mar 2017
PROBLEM MOTIVATION
The explorationexploitation dilemma arises in the following three problems of active learning of Gaussian processes:
 NonStationary Gaussian Processes. A key challenge of environmental sensing and monitoring is that of sensing, modeling, and predicting complex urban and natural environmental phenomena, which are typically characterized by spatially correlated measurements. To tackle this challenge, recent research efforts in the robotics community have focused on developing multirobot active sensing (MAS) algorithms: Their objective is to coordinate the exploration of a team of mobile robots to actively
gather the most informative observations for predicting a spatially varying phenomenon of interest while being subject to resource cost constraints (e.g., number of deployed robots, energy consumption, mission time). To achieve this, a number of MAS algorithms have modeled the phenomenon as a Gaussian process (GP), which allows its spatial correlation structure to be formally characterized and its predictive uncertainty to be formally quantified (e.g., based on meansquared error, entropy, or mutual information criterion) and subsequently exploited for directing the robots to explore its highly uncertain areas. In order not to incur high computational expense, these algorithms have assumed the spatial correlation structure to be known (or estimated crudely using sparse prior data) and stationary (i.e., degree of smoothness in the spatial variation of the measurements is the same across the entire phenomenon), properties of which are often violated in realworld environmental sensing applications and limited to smallscale phenomena.
In practice, the spatial correlation structure of possibly largescale environmental phenomena is usually not known and nonstationary (i.e., separate areas of a phenomenon exhibit different local degrees of smoothness in the spatial variation of the measurements). For example, in some ocean phenomena (e.g., temperature, salinity, sea surface height), their measurements far offshore are more smoothly varying (i.e., more spatially correlated) in the crossshore direction than nearshore. Urban traffic networks also display nonstationary phenomena (e.g., traffic speeds, taxi demands), which pose important considerations to traffic routing and signal control.
Existing MAS algorithms can still be used for sampling a nonstationary phenomenon by assuming, albeit incorrectly, its spatial correlation structure to be known and stationary in order to preserve time efficiency. So, though they can gather the most informative observations under an assumed stationary correlation structure, they will perform suboptimally with respect to the true nonstationary correlation structure.
A more desirable MAS algorithm should instead be designed to consider the informativeness of its selected observations for both estimating the unknown spatial correlation structure of a phenomenon (i.e., exploration) as well as predicting the phenomenon given the true correlation structure (i.e., exploitation). According to previous geostatistical studies, the most informative observations that are gathered for achieving the former active sensing criterion are not necessarily as informative for satisfying the latter. This raises a fundamental issue faced by active sensing: How can a MAS algorithm trade off between these two possibly conflicting criteria?
 Nonmyopic Active Learning. Active learning has become an increasingly important focal theme in many environmental sensing and monitoring applications (e.g., precision agriculture, mineral prospecting, monitoring of ocean and freshwater phenomena like harmful algal blooms, forest ecosystems, or pollution) where a highresolution in situ sampling of the spatial phenomenon of interest is impractical due to prohibitively costly sampling budget requirements (e.g., number of deployed sensors, energy consumption, mission time): For such applications, it is thus desirable to select and gather the most informative observations/data for modeling and predicting the spatially varying phenomenon subject to some budget constraints, which is the goal of active learning and also known as the active sensing problem.
To elaborate, solving the active sensing problem amounts to deriving an optimal sequential policy that plans/decides the most informative locations to be observed for minimizing the predictive uncertainty of the unobserved areas of a phenomenon given a sampling budget. To achieve this, many existing active sensing algorithms have modeled the phenomenon as a Gaussian process (GP), which allows its spatial correlation structure to be formally characterized and its predictive uncertainty to be formally quantified (e.g., based on meansquared error, entropy, or mutual information criterion). However, they have assumed the spatial correlation structure (specifically, the parameters defining it) to be known, which is often violated in realworld applications, or estimated crudely using sparse prior data. So, though they aim to select sampling locations that are optimal with respect to the assumed or estimated parameters, these locations tend to be suboptimal with respect to the true parameters, thus degrading the predictive performance of the learned GP model.
In practice, the spatial correlation structure of a phenomenon is usually not known. Then, the predictive performance of the GP modeling the phenomenon depends on how informative the gathered observations/data are for both parameter estimation as well as spatial prediction given the true parameters. Interestingly, as revealed in previous geostatistical studies, policies that are efficient for parameter estimation are not necessarily efficient for spatial prediction with respect to the true model. Thus, the active sensing problem involves a potential tradeoff between sampling the most informative locations for spatial prediction given the current, possibly incomplete knowledge of the model parameters (i.e., exploitation) vs. observing locations that gain more information about the parameters (i.e., exploration): How then does an active sensing algorithm trade off between these two possibly conflicting sampling objectives?
To tackle this question, one principled approach is to frame active sensing as a sequential decision problem that jointly and naturally optimizes the above explorationexploitation tradeoff while maintaining a Bayesian belief over the model parameters. This intuitively means a policy that biases towards observing informative locations for spatial prediction given the current model prior may be penalized if it entails a highly dispersed posterior over the model parameters. So, the resulting induced policy is guaranteed to be optimal in the expected active sensing performance. Unfortunately, such a nonmyopic Bayesoptimal policy cannot be derived exactly due to an uncountable set of candidate observations and unknown model parameters. As a result, most existing works have circumvented the tradeoff by resorting to the use of myopic/greedy (hence, suboptimal) policies. To the best of our knowledge, the only notable nonmyopic active sensing algorithm for GPs advocates tackling exploration and exploitation separately, instead of jointly and naturally optimizing their tradeoff, to sidestep the difficulty of solving the Bayesian sequential decision problem. Specifically, it performs a probably approximately correct (PAC)style exploration until it can verify that the performance loss of greedy exploitation lies within a userspecified threshold. But, such an algorithm is suboptimal in the presence of budget con straints due to the following limitations: (a) It is unclear how an optimal threshold for exploration can be determined given a sampling budget, and (b) even if such a threshold is available, the PACstyle exploration is typically designed to satisfy a worstcase sample complexity rather than to be optimal in the expected active sensing performance, thus resulting in an overlyaggressive exploration.
 MultiOutput Gaussian Processes. For many budgetconstrained environmental sensing and monitoring applications in the real world, active learning/sensing is an attractive, frugal alternative to passive highresolution (hence, prohibitively costly) sampling of the spatially varying target phenomenon of interest. Different from the latter, active learning aims to select and gather the most informative observations for modeling and predicting the spatially varying phenomenon given some sampling budget constraints (e.g., quantity of deployed sensors, energy consumption, mission time).
In practice, the target phenomenon often coexists and correlates well with some auxiliary type(s) of phenomena whose measurements may be more spatially correlated, less noisy (e.g., due to higherquality sensors), and/or less tedious to sample (e.g., due to greater availability/quantity, higher sampling rate, and/or lower sampling cost of deployed sensors of these type(s)) and can consequently be exploited for improving its prediction. For example, to monitor soil pollution by some heavy metal (e.g., Cadmium), its complex and timeconsuming extraction from soil samples can be alleviated by supplementing its prediction with correlated auxiliary types of soil measurements (e.g., pH) that are easier to sample. Similarly, to monitor algal bloom in the coastal ocean, plankton abundance correlates well with auxiliary types of ocean measurements (e.g., chlorophyll a, temperature, and salinity) that can be sampled more readily. Other examples of realworld applications include remote sensing, traffic monitoring, monitoring of groundwater and indoor environmental quality, and precision agriculture, among others. All of the above practical examples motivate the need to design and develop an active learning algorithm that selects not just the most informative sampling locations to be observed but also the types of measurements (i.e., target and/or auxiliary) at each selected location for minimizing the predictive uncertainty of unobserved areas of a target phenomenon given a sampling budget, which is the focus of our work here.
To achieve this, we model all types of coexisting phenomena (i.e., target and auxiliary) jointly as a multioutput Gaussian process (MOGP), which allows the spatial correlation structure of each type of phenomenon and the crosscorrelation structure between different types of phenomena to be formally characterized. More importantly, unlike the nonprobabilistic multivariate regression methods, the probabilistic MOGP regression model allows the predictive uncertainty of the target phenomenon (as well as the auxiliary phenomena) to be formally quantified (e.g., based on entropy or mutual information criterion) and consequently exploited for deriving the active learning criterion.
PROPOSED METHODOLOGY
 NonStationary Gaussian Processes. To address the first problem, our first work presents a decentralized multirobot active sensing (DECMAS) algorithm that can efficiently coordinate the exploration of multiple robots to jointly optimize the above tradeoff for sampling unknown, nonstationary environmental phenomena. Our DECMAS algorithm models a nonstationary phenomenon as a Dirichlet process mixture of Gaussian processes (DPMGPs): Using the gathered observations, DPMGPs can learn to automatically partition the phenomenon into separate local areas, each of which comprises measurements that vary according to a stationary spatial correlation structure and can thus be modeled by a locally stationary GP. The main contributions of our work here are novel in demonstrating how DPMGPs and its structural properties can be exploited to (a) formalize an active sensing criterion that trades off between gathering the most informative observations for estimating the unknown partition (i.e., a key component of the nonstationary correlation structure) vs. that for predicting the phenomenon given the current, possibly imprecise estimate of the partition, and (b) support effective and efficient decentralized coordination. We also provide a theoretical performance guarantee for DECMAS and analyze its time complexity. Finally, we empirically demonstrate using two realworld datasets that DECMAS outperforms the stateoftheart MAS algorithms.
 Nonmyopic Active Learning. To tackle the second problem, our second work presents an efficient decisiontheoretic planning approach to nonmyopic active sensing/learning that can still preserve and exploit the principled Bayesian sequential decision problem framework for jointly and naturally optimizing the explorationexploitation tradeoff and consequently does not incur the limitations of the algorithm of Krause & Guestrin (2007). In particular, although the exact Bayesoptimal policy to the active sensing problem cannot be derived, we show that it is in fact possible to solve for a nonmyopic ϵBayesoptimal active learning (ϵBAL) policy given a userdefined bound ϵ, which is the main contribution of our work here. In other words, our proposed ϵBAL policy can approximate the optimal expected active sensing performance arbitrarily closely (i.e., within an arbitrary loss bound ϵ). In contrast, the algorithm of Krause & Guestrin (2007) can only yield a suboptimal performance bound. To meet the realtime requirement in timecritical applications, we then propose an asymptotically ϵoptimal, branchandbound anytime algorithm based on ϵBAL with performance guarantee. We empirically demonstrate using both synthetic and realworld datasets that, with limited budget, our proposed approach outperforms stateoftheart algorithms.
 MultiOutput Gaussian Processes. To solve the third problem,
our third work is the first to present an efficient algorithm for active learning of a MOGP model. We consider utilizing the entropy criterion to measure the predictive uncertainty of a target phenomenon, which is widely used for active learning of a singleoutput GP model. Unfortunately, for the MOGP model, such a criterion scales poorly in the number of candidate sampling locations of the target phenomenon and even more so in the number of selected observations (i.e., sampling budget) when optimized. To resolve this scalability issue, we first exploit a structure common to a unifying framework of sparse MOGP models for deriving a novel active learning criterion.
Our novel active learning criterion exhibits an interesting explorationexploitation tradeoff between
selecting locations with the most uncertain measurements of the target phenomenon to be observed given the latent structure of the sparse MOGP model (i.e., exploitation) vs. selecting locations to be observed (i.e., possibly of auxiliary types of phenomena) so as to rely less on measurements at the remaining unobserved locations (i.e., won't be sampled) of the target phenomenon to infer the latent model structure (i.e., exploration).
Then, we define a relaxed notion of submodularity called ϵsubmodularity and exploit the ϵsubmodularity property of our new criterion for devising a polynomialtime approximation algorithm that guarantees a constantfactor approximation of that achieved by the optimal set of selected observations. We empirically evaluate the performance of our proposed algorithm using three realworld datasets.
PUBLICATIONS
 An InformationTheoretic Framework for Unifying Active Learning Problems.
Quoc Phong Nguyen, Kian Hsiang Low & Patrick Jaillet.
In Proceedings of the 35th AAAI Conference on Artificial Intelligence (AAAI21), pages 91269134, Feb 29, 2021.
21.4% acceptance rate
Abstract. This paper presents an informationtheoretic framework for unifying active learning problems: level set estimation (LSE), Bayesian optimization (BO), and their generalized variant. We first introduce a novel active learning criterion that subsumes an existing LSE algorithm and achieves stateoftheart performance in LSE problems with a continuous input domain. Then, by exploiting the relationship between LSE and BO, we design a competitive informationtheoretic acquisition function for BO that has interesting connections to upper confidence bound and maxvalue entropy search (MES). The latter connection reveals a drawback of MES which has important implications on not only MES but also on other MESbased acquisition functions. Finally, our unifying informationtheoretic framework can be applied to solve a generalized problem of LSE and BO involving multiple level sets in a dataefficient manner. We empirically evaluate the performance of our proposed algorithms using synthetic benchmark functions, a realworld dataset, and in hyperparameter tuning of machine learning models.
 DataEfficient Machine Learning with Multiple Output Types and High Input Dimensions.
Yehong Zhang^{}.
Ph.D. Thesis, Department of Computer Science, National University of Singapore, Dec 2017.
Abstract.
Recent research works in machine learning (ML) have focused on learning some target variables of interest to achieve competitive (or stateoftheart) predictive performance in less time but without requiring large quantities of data, which is known as dataefficient ML. This thesis focuses on two important dataefficient ML approaches: active learning (AL) and Bayesian optimization (BO) which, instead of learning passively from a given small set of data, need to select and gather the most informative observations for learning the target variables of interest more accurately given some budget constraints. To advance the stateoftheart of dataefficient ML, novel generalizations of AL and BO algorithms are proposed in this thesis for addressing the issues arising from multiple output types and high input dimensions which are the practical settings in many realworld applications.
In particular, this thesis aims to (a) exploit the auxiliary types of outputs which usually coexist and correlate well with the target output types, and more importantly, are less noisy and/or less tedious to sample for improving the learning performance of the target output type in both AL and BO algorithms and (b) scale up the stateoftheart BO algorithm to high input dimensions. To achieve this, the specific data with multiple output types or high input dimensions is represented using some form of Gaussian process (GP)based probabilistic regression models which allow the predictive uncertainty of the outputs to be formally quantified and consequently exploited for developing efficient AL and BO algorithms.
To achieve above objectives, an AL algorithm of multioutput GP (MOGP) is first developed for minimizing the predictive uncertainty (i.e., posterior joint entropy) of the target output type. In contrast to existing works, our AL problems involve selecting not just the most informative sampling inputs to be observed but also the types of outputs at each selected input for improving the learning performance of only the target output type given a sampling budget. Unfortunately, such an entropy criterion scales poorly in the numbers of candidate sampling inputs and selected observations when optimized. To resolve this issue, we exploit a structure common to sparse MOGP models for deriving a novel AL criterion. Furthermore, we exploit a relaxed form of submodularity property of our new criterion for devising a polynomialtime approximation algorithm that guarantees a constantfactor approximation of that achieved by the optimal set of selected observations. Empirical evaluation on realworld datasets shows that our proposed approach outperforms existing algorithms for AL of MOGP and singleoutput GP models.
Secondly, to boost the BO performance by exploiting the cheaper and/or less noisy observations of some auxiliary functions with varying fidelities, we proposed a novel generalization of predictive entropy search (PES) for multifidelity BO called multifidelity PES (MFPES). In contrast to existing multifidelity BO algorithms, our proposed MFPES algorithm can naturally trade off between exploitation vs. exploration over the target and auxiliary functions with varying fidelities without needing to manually tune any such parameters. To achieve this, we model the unknown target and auxiliary functions jointly as a convolved MOGP (CMOGP) whose convolutional structure is exploited to formally characterize the fidelity of each auxiliary function through its crosscorrelation with the target function. Although the exact acquisition function of MFPES cannot be computed in closed form, we show that it is in fact possible to derive an efficient approximation of MFPES via a novel multioutput random features approximation of the CMOGP model whose crosscorrelation (i.e., multifidelity) structure between the target and auxiliary functions can be exploited for improving the belief of the global target maximizer using the observations from evaluating these functions. Practical constraints are proposed to relate the global target maximizer to that of auxiliary functions. Empirical evaluation on synthetic and realworld experiments shows that MFPES outperforms the stateoftheart multifidelity BO algorithms.
Lastly, to improve the BO performance in realworld applications with high input
dimensions (e.g., computer vision, biology), we generalize PES for highdimensional BO by exploiting an additive structure of the target function. New practical constraints are proposed and approximated efficiently such that the proposed acquisition function of additive PES (addPES) can be optimized independently for each local and lowdimensional input component. The empirical results show that our addPES considerably improves the performance of the stateoftheart highdimensional BO algorithms by using a simple and common setting for optimizing different tested functions with varying input dimensions, which makes it a superior alternative to existing highdimensional BO algorithms.
 Exploiting Decentralized MultiAgent Coordination for LargeScale Machine Learning Problems.
Ruofei Ouyang.
Ph.D. Thesis, Department of Computer Science, National University of Singapore, Dec 2016.
Abstract.
Nowadays, the scale of machine learning problems becomes much larger than before. It raises a huge demand in distributed perception and distributed computation. A multiagent system provides exceptional scalability for problems like active sensing and data fusion. However, many rich characteristics of largescale machine learning problems have not been addressed yet such as large input domain, nonstationarity, and high dimensionality. This thesis identifies the challenges related to these characteristics from multiagent perspective. By exploiting the correlation structure of data in largescale problems, we propose multiagent coordination schemes that can improve the scalability of the machine learning models while preserving the computation accuracy. To elaborate, the machine learning problems we are solving with multiagent coordination techniques are:
 Gaussian process regression. To perform distributed regression on a largescale environmental phenomenon, data compression is often required due to the communication costs. Currently, decentralized data fusion methods encapsulate the data into local summaries based on a fixed support set. However in a largescale field, this fixed support set, acting as a centralized component in the decentralized system, cannot approximate the correlation structure of the entire phenomenon well. It leads to evident losses in data summarization. Consequently, the regression performance will be significantly reduced.
In order to approximate the correlation structure accurately, we propose an agentcentric support set to allow every agent in the data fusion system to choose a possibly different support set and dynamically switch to another one during execution for encapsulating its own data into a local summary which, perhaps surprisingly, can still be assimilated with the other agents’ local summaries into a globally consistent summary. Together with an information sharing mechanism we designed, the new decentralized data fusion methods with agentcentric support set can be applied to regression problems on a much larger environmental phenomenon with high performance.
 Active learning. In the context of environmental sensing, active learning/active sensing is a process of taking observations to minimize the uncertainty in an environmental field. The uncertainty is quantified based on the correlation structure of the phenomenon which is traditionally assumed to be stationary for computational sake. In a largescale environmental field, this stationary assumption is often violated. Therefore, existing active sensing algorithms perform suboptimally for a nonstationary environmental phenomenon.
To the best of our knowledge, our decentralized multirobot active sensing (DECMAS) algorithm is the first work to address nonstationarity issue in the context of active sensing. The uncertainty in the phenomenon is quantified based on the nonstationary correlation structure estimated by Dirichlet process mixture of Gaussian processes. Further, our DECMAS algorithm can efficiently coordinate the exploration of multiple robots to automatically tradeoff between learning the unknown, nonstationary correlation structure and minimizing the uncertainty of the environmental phenomenon. It enables multiagent active sensing techniques to be applied to a largescale nonstationary environmental phenomenon.
 Bayesian optimization. Optimizing an unknown objective function is challenging for traditional optimization methods. Alternatively, in this situation, people use Bayesian optimization which is a modern optimization technique that can optimize a function by only utilizing the observation information (input and output values) collected through simulations. When the input dimension of the function is low, a few simulated observations can generate good result already. However, for high dimensional function, a huge number of observations are required which is impractical when the simulation consumes lots of time and resources.
Fortunately, many high dimensional problems have sparse correlation structure. Our ANOVADCOP work can decompose the correlation structure in the original highdimensional problem into many correlation structures of subsets of dimensions based on ANOVA kernel function. It significantly reduces the size of input space into a collection of lowerdimensional subspaces. Additionally, we reformulate the Bayesian optimization problem as a decentralized constrained optimization problem (DCOP) that can be efficiently solved by multiagent coordination techniques so that it can scale up to problems with hundreds of dimensions.
 NearOptimal Active Learning of MultiOutput Gaussian Processes.
Yehong Zhang, Trong Nghia Hoang, Kian Hsiang Low & Mohan Kankanhalli.
In Proceedings of the 30th AAAI Conference on Artificial Intelligence (AAAI16), pages 23512357, Phoenix, AZ, Feb 1217, 2016.
25.75% acceptance rate
Abstract. This paper addresses the problem of active learning of a multioutput Gaussian process (MOGP) model representing multiple types of coexisting correlated environmental phenomena. In contrast to existing works, our active learning problem involves selecting not just the most informative sampling locations to be observed but also the types of measurements at each selected location for minimizing the predictive uncertainty (i.e., posterior joint entropy) of a target phenomenon of interest given a sampling budget. Unfortunately, such an entropy criterion scales poorly in the numbers of candidate sampling locations and selected observations when optimized. To resolve this issue, we first exploit a structure common to sparse MOGP models for deriving a novel active learning criterion. Then, we exploit a relaxed form of submodularity property of our new criterion for devising a polynomialtime approximation algorithm that guarantees a constantfactor approximation of that achieved by the optimal set of selected observations. Empirical evaluation on realworld datasets shows that our proposed approach outperforms existing algorithms for active learning of MOGP and singleoutput GP models.
 New Advances on Bayesian and DecisionTheoretic Approaches for Interactive Machine Learning.
Trong Nghia Hoang.
Ph.D. Thesis, Department of Computer Science, National University of Singapore, Feb 2015.
Abstract.
The explorationexploitation tradeoff is a fundamental dilemma in many interactive learning scenarios which include both aspects of reinforcement learning (RL) and active learning (AL): An autonomous agent, situated in an unknown environment, has to actively extract knowledge from the environment by taking actions (or conducting experiments) based on its previously collected information to make accurate predictions or to optimize some utility functions. Thus, to make the most effective use of their resourceconstrained budget (e.g., processing time, experimentation cost), the agent must choose carefully between (a) exploiting options (e.g., actions, experiments) which are recommended by its current, possibly incomplete model of the environment, and (b) exploring the other ostensibly suboptimal choices to gather more information.
For example, an RL agent has to face a dilemma between (a) exploiting the mostrewarding action according to the current statistical model of the environment at the risk of running into catastrophic situations if the model is not accurate, and (b) exploring a suboptimal action to gather more information so as to improve the model's accuracy at the potential price of losing the shortterm reward. Similarly, an AL algorithm/agent has to consider between (a) conducting the most informative experiments according to its current estimation of the environment model's parameters (i.e., exploitation), and (b) running experiments that help improving the estimation accuracy of these parameters (i.e., exploration).
More often, learning strategies that ignore exploration will likely exhibit suboptimal performance due to their imperfect knowledge while, conversely, those that entirely focus on exploration might suffer the cost of learning without benefitting from it. Therefore, a good explorationexploitation tradeoff is critical to the success of those interactive learning agents: In order to perform well, they must strike the right balance between these two conflicting objectives. Unfortunately, while this tradeoff has been wellrecognized since the early days of RL, the studies of explorationexploitation have mostly been developed for theoretical settings in the respective field of RL and, perhaps surprisingly, glossed over in the existing AL literature. From a practical point of view, we see three limiting factors:
 Previous works addressing the explorationexploitation tradeoff in RL have largely focused on simple choices of the environment model and consequently, are not practical enough to accommodate realworld applications that have far more complicated environment structures. In fact, we find that most recent advances in Bayesian reinforcement learning (BRL) have only been able to analytically trade off between exploration and exploitation under a simple choice of models such as FlatDirichletMultinomial (FDM) whose independence and modeling assumptions do not hold for many realworld applications.
 Nearly all of the notable works in the AL literature primarily advocate the use of greedy/myopic algorithms whose rates of convergence (i.e., the number of experiments required by the learning algorithm to achieve a desired performance in the worst case) are provably minimax optimal for simple classes of learning tasks (e.g., threshold learning). While these results have greatly ad vanced our understanding about the limit of myopic AL in worstcase scenarios, significantly less is presently known about whether it is possible to devise nonmyopic AL strategies which optimize the explorationexploitation tradeoff to achieve the best expected performance in budgeted learning scenarios.
 The issue of scalability of the existing predictive models (e.g., Gaussian processes) used in AL has generally been underrated since the majority of literature considers smallscale environments which only consist of a few thousand candidate experiments to be selected by singlemode AL algorithms one at a time prior to retraining the model. In contrast, largescale environments usually have a massive set of million candidate experiments among which tens or hundreds of thousands should be actively selected for learning. For such dataintensive problems, it is often more costeffective to consider batchmode AL algorithms which select and conduct multiple experiments in parallel at each stage to collect observations in batch. Retraining the predictive model after incorporating each batch of observations then becomes a computational bottleneck as the collected dataset at each stage quickly grows up to tens or even hundreds of thousand data points.
This thesis outlines some recent progresses that we have been able to make while working toward satisfactory answers to the above challenges, along with practical algorithms that achieve them:
 In particular, in order to put BRL into practice for more complicated and practical problems, we propose a novel framework called Interactive Bayesian Reinforcement Learning (IBRL) to integrate the general class of parametric models and model priors, thus allowing the practitioners' domain knowledge to be exploited to produce a finegrained and compact representation of the environment as often required in many realworld applications. Interestingly, we show how the nonmyopic Bayesoptimal policy can be derived analytically by solving IBRL exactly and propose an approximation algorithm to compute it efficiently in polynomial time. Our empirical studies show that the proposed approach performs competitively with the existing stateoftheart algorithms.
 Then, to establish a theoretical foundation for the explorationexploitation tradeoff in singlemode active learning scenarios with resourceconstrained budgets, we present a novel ϵBayesoptimal DecisionTheoretic Active Learning (ϵBAL) framework which advocates the use of differential entropy as a performance measure and consequently, derives a learning policy that can approximate the optimal expected performance arbitrarily closely (i.e., within an arbitrary loss bound ϵ). To meet the realtime requirement in timecritical applications, we then propose an asymptotically ϵoptimal, branchandbound anytime algorithm based on ϵBAL with performance guarantees. In practice, we empirically demonstrate with both synthetic and realworld datasets that the proposed approach outperforms the stateoftheart algorithms in budgeted scenarios.
 Lastly, to facilitate the future developments of largescale, nonmyopic AL applications, we further introduce a highly scalable family of anytime predictive models for AL which provably converge toward a wellknown class of sparse Gaussian processes (SGPs). Unlike the existing predictive models of AL which cannot be updated incrementally and are only capable of processing middlesized datasets (i.e., a few thousands of data points), our proposed models can process massive datasets in an anytime fashion, thus providing a principled tradeoff between the processing time and the predictive accuracy. The efficiency of our framework is then demonstrated empirically on a variety of largescale realworld datasets which contains hundreds of thousand data points.
 Nonmyopic ϵBayesOptimal Active Learning of Gaussian Processes.
Trong Nghia Hoang, Kian Hsiang Low, Patrick Jaillet and Mohan Kankanhalli.
In Proceedings of the 31st International Conference on Machine Learning (ICML14), pages 739747, Beijing, China, Jun 2126, 2014.
22.4% acceptance rate (cycle 2)
Also appeared in
RSS14 Workshop on NonParametric Learning in Robotics, Berkeley, CA, Jul 12, 2014.
Abstract. A fundamental issue in active learning of Gaussian processes is that of the explorationexploitation tradeoff.
This paper presents a novel nonmyopic ϵBayesoptimal active learning (ϵBAL) approach that jointly and naturally optimizes the tradeoff.
In contrast, existing works have primarily developed myopic/greedy algorithms or performed exploration and exploitation separately.
To perform active learning in real time, we then propose an anytime algorithm based on ϵBAL with performance guarantee and empirically demonstrate using synthetic and realworld datasets that, with limited budget, it outperforms the stateoftheart algorithms.
 Active Learning is Planning: Nonmyopic ϵBayesOptimal Active Learning of Gaussian Processes.
Trong Nghia Hoang, Kian Hsiang Low, Patrick Jaillet and Mohan Kankanhalli.
In T. Calders, F. Esposito, E. Hüllermeier, R. Meo, editors, Machine Learning and Knowledge Discovery in Databases  European Conference, ECML/PKDD14 Nectar (New Scientific and Technical Advances in Research) Track, Part III, LNCS 8726, pages 494498, Springer Berlin Heidelberg, Nancy, France, Sep 1519, 2014.
Abstract. A fundamental issue in active learning of Gaussian processes is that of the explorationexploitation tradeoff. This paper presents a novel nonmyopic ϵBayesoptimal active learning (ϵBAL) approach that jointly optimizes the tradeoff. In contrast, existing works have primarily developed greedy algorithms or performed exploration and exploitation separately. To perform active learning in real time, we then propose an anytime algorithm based on ϵBAL with performance guarantee and empirically demonstrate using a realworld dataset that, with limited budget, it outperforms the stateoftheart algorithms.
 MultiRobot Active Sensing of NonStationary Gaussian ProcessBased Environmental Phenomena.
Ruofei Ouyang, Kian Hsiang Low, Jie Chen & Patrick Jaillet.
In Proceedings of the
13th International Conference on Autonomous Agents and MultiAgent Systems (AAMAS14), pages 573580, Paris, France, May 59, 2014.
23.8% acceptance rate
Also appeared in
RSS14 Workshop on NonParametric Learning in Robotics, Berkeley, CA, Jul 12, 2014.
Abstract. A key challenge of environmental sensing and monitoring is that of sensing, modeling, and predicting largescale, spatially correlated environmental phenomena, especially when they are unknown and nonstationary.
This paper presents a decentralized multirobot active sensing (DECMAS) algorithm that can efficiently coordinate the exploration of multiple robots to gather the most informative observations for predicting an unknown, nonstationary phenomenon.
By modeling the phenomenon using a Dirichlet process mixture of Gaussian processes (DPMGPs), our work here is novel in demonstrating how DPMGPs and its structural properties can be exploited to (a) formalize an active sensing criterion that trades off between gathering the most informative observations for estimating the unknown, nonstationary spatial correlation structure vs. that for predicting the phenomenon given the current, imprecise estimate of the correlation structure, and (b) support efficient decentralized coordination.
We also provide a theoretical performance guarantee for DECMAS and analyze its time complexity.
We empirically demonstrate using two realworld datasets that DECMAS outperforms stateoftheart MAS algorithms.
ONLINE AND ANYTIME SPARSE GAUSSIAN PROCESSES FOR BIG DATA
PROJECT DURATION : Aug 2013  Present
PROJECT AFFILIATION

SingaporeMIT Alliance for Research and Technology (SMART) Future Urban Mobility (FM) IRG (Collaborators: Emilio Frazzoli, MIT; Daniela Rus, MIT)

SensorEnhanced Social Media (SeSaMe) Centre (Collaborator: Mohan Kankanhalli)
PROJECT FUNDING
 RIE2020 AME Programmatic Fund : Learning with Less Data, SGD $1,218,600, Apr 2021  Mar 2024
 MOE AcRF Tier 1 Reimagine Research Scheme Funding : Scalable AI Phenome Platform towards FastForward Plant Breeding (Machine Learning),
SGD $348,600, Mar 2021  Mar 2024
 RIE2020 AME IAFPP : High Performance Precision Agriculture (HiPPA) System, S$1,197,960, Mar 2020  Feb 2024
 Research Collaboration Agreement with InfoCommunications Media Development Authority (IMDA) : Robust and Scalable Computer Vision for Scene Understanding, SGD $260,000, Oct 2020  Oct 2021
 MOE AcRF Tier 2 Grant : Scaling up Gaussian Process Predictive Models for Big Data, SGD $737,461, Jul 2017  Jul 2020
 SMART Subaward Agreements  FM IRG :
Autonomy in MobilityOnDemand Systems,
SGD $1,348,638.22, Aug 2010  Dec 2015
 Research Collaboration Agreements with Panasonic R&D Center Singapore : Sonar Data Fusion Algorithm for Object Distance Estimation, SGD $84,230.40, Feb 2016  Jul 2016, Dec 2016  Jul 2017
PROBLEM MOTIVATION
A Gaussian process regression (GPR) model is a Bayesian nonparametric model for performing nonlinear regression that provides a Gaussian predictive distribution with formal measures of predictive uncertainty. The expressivity of a fullrank GPR (FGPR) model, however, comes at a cost of cubic time in the size of the data, thus rendering it computationally impractical for training with massive datasets. To improve its scalability, a number of sparse GPR (SGPR) models
exploiting lowrank approximate representations have been proposed, many of which share a similar structural assumption of conditional independence (albeit of varying degrees) based on the notion of inducing variables and consequently incur only linear time in the data size. The work of QuinoneroCandela & Rasmussen (2005) has in fact presented a unifying view of such SGPR models, which include the subset of regressors (SoR), deterministic training conditional (DTC), fully independent training conditional (FITC), fully independent conditional (FIC), partially independent training conditional (PITC), and partially independent conditional (PIC) approximations.
To scale up these SGPR models further for performing realtime predictions necessary in many timecritical applications and decision support systems (e.g., ocean sensing, traffic monitoring), the work of Gal et al. (2014) has parallelized DTC while that of Chen et al. (2013) has parallelized FITC, FIC, PITC, and PIC to be run on multiple machines. The recent work of Low et al. (2015) has produced a spectrum of SGPR models with PIC and FGPR at the two extremes that are also amenable to parallelization on multiple machines. Ideally, these parallel SGPR models can reduce the incurred time of their centralized counterparts by a factor close to the number of machines. In practice, since the number of machines is limited due to budget constraints, their incurred time will still grow with an increasing size of data. Like their centralized counterparts, they can be trained using all the data.
When the data is expected to stream in over a (possibly indefinitely) long time, it is also computationally impractical to repeatedly use these existing offline sparse GP approximation methods or even the online GP model (i.e., quadratic time in the data size) for training at each time step.
PROPOSED METHODOLOGY
A more affordable alternative is to instead train a SGPR model in either an (1) online or (2) anytime fashion with a small, randomly sampled subset of the data at each iteration, which requires only a single machine:
 Our first work presents a novel online sparse GP approximation method that, in contrast to existing works mentioned above, is capable of achieving constant time and memory (i.e., independent of the size of the data/observations) per time step. We provide a theoretical guarantee on its predictive performance to be equivalent to that of the offline sparse PITC approximation method. Our proposed method generalizes the sparse online GP model of Csato & Opper (2002) by relaxing its conditional independence assumption significantly, hence potentially improving the predictive performance. We empirically demonstrate the practical feasibility of using our generalized online sparse GP approximation method through a realworld persistent mobile robot localization experiment.
 To the best of our knowledge, the only notable anytime SGPR model exploits a result of Titsias (2009) that DTC can alternatively be obtained using variational inference by minimizing the KullbackLeibler (KL) distance between the variational approximation and the GP posterior distribution of some latent variables given the data, from which a stochastic natural gradient ascent (SNGA) method can be derived to achieve an asymptotic convergence of its predictive performance to that of DTC while incurring constant time per iteration.
This anytime variant of DTC promises a huge speedup if the number of sampled subsets of data needed for convergence is much smaller than the total number of possible disjoint subsets that can be formed and sampled from all the data. But, it can be observed in our experiments that DTC often does not predict as well as the other SGPR models (except SoR) encompassed by the unifying view of QuinoneroCandela & Rasmussen (2005) because it imposes the most restrictive structural assumption. This motivates us to consider the possibility of constructing an anytime variant of any SGPR model of our choice whose derived SNGA method can achieve an asymptotic convergence of its predictive performance to that of the chosen SGPR model while preserving constant time per iteration.
However, no alternative formulation based on variational inference exists for any SGPR model other than DTC in order to derive such a SNGA method.
To address the above challenge, our second work presents a novel unifying framework of anytime SGPR models that can produce good predictive performance fast and improve their predictive performance over time. Our proposed unifying framework, perhaps surprisingly, reverses the variational inference procedure to theoretically construct a nontrivial, concave functional (i.e., of distributions) that is maximized at the predictive distribution of any SGPR model of our choice. Consequently, a SNGA method can be derived that involves iteratively following the stochastic natural gradient of the functional to improve its estimate of the predictive distribution of the chosen SGPR model and is guaranteed to achieve asymptotic convergence to it.
Interestingly, we show that if the predictive distribution of the chosen SGPR model satisfies certain decomposability conditions (e.g., DTC, FITC, PIC), then the stochastic natural gradient is an unbiased estimator of the exact natural gradient and can be computed in constant time (i.e., independent of data size) at each iteration. We empirically evaluate the tradeoff between the predictive performance vs. time efficiency of the anytime SGPR models spanned by our unifying framework (i.e., including stateoftheart anytime variant of DTC) on two realworld millionsized datasets.
PUBLICATIONS
 Convolutional Normalizing Flows for Deep Gaussian Processes.
Haibin Yu, Dapeng Liu, Kian Hsiang Low & Patrick Jaillet.
In Proceedings of the International Joint Conference on Neural Networks (IJCNN'21), Jul 1822, 2021.
Abstract. Deep Gaussian processes (DGPs), a hierarchical composition of GP models, have successfully boosted the expressive power of their singlelayer counterpart. However, it is impossible to perform exact inference in DGPs, which has motivated the recent development of variational inferencebased methods. Unfortunately, either these methods yield a biased posterior belief or it is difficult to evaluate their convergence. This paper introduces a new approach for specifying flexible, arbitrarily complex, and scalable approximate posterior distributions. The posterior distribution is constructed through a normalizing flow (NF) which transforms a simple initial probability into a more complex one through a sequence of invertible transformations. Moreover, a novel convolutional normalizing flow (CNF) is developed to improve the time efficiency and capture dependency between layers. Empirical evaluation shows that CNF DGP outperforms the stateoftheart approximation methods for DGPs.
 Scalable Variational Bayesian Kernel Selection for Sparse Gaussian Process Regression.
Tong Teng, Jie Chen, Yehong Zhang & Kian Hsiang Low.
In Proceedings of the 34th AAAI Conference on Artificial Intelligence (AAAI20), pages 59976004, New York, NY, Feb 712, 2020.
20.6% acceptance rate
Abstract. This paper presents a variational Bayesian kernel selection (VBKS) algorithm for sparse Gaussian process regression (SGPR) models. In contrast to existing GP kernel selection algorithms that aim to select only one kernel with the highest model evidence, our proposed VBKS algorithm considers the kernel as a random variable and learns its belief from data such that the uncertainty of the kernel can be interpreted and exploited to avoid overconfident GP predictions. To achieve this, we represent the probabilistic kernel as an additional variational variable in a variational inference (VI) framework for SGPR models where its posterior belief is learned together with that of the other variational variables (i.e., inducing variables and kernel hyperparameters). In particular, we transform the discrete kernel belief into a continuous parametric distribution via reparameterization in order to apply VI. Though it is computationally challenging to jointly optimize a large number of hyperparameters due to many kernels being evaluated simultaneously by our VBKS algorithm, we show that the variational lower bound of the logmarginal likelihood can be decomposed into an additive form such that each additive term depends only on a disjoint subset of the variational variables and can thus be optimized independently. Stochastic optimization is then used to maximize the variational lower bound by iteratively improving the variational approximation of the exact posterior belief via stochastic gradient ascent, which incurs constant time per iteration and hence scales to big data. We empirically evaluate the performance of our VBKS algorithm on synthetic and massive realworld datasets.
 New Advances in Bayesian Inference for Gaussian Process and Deep Gaussian Process Models.
Haibin Yu^{}.
Ph.D. Thesis, Department of Computer Science, National University of Singapore, May 2020.
Abstract.
Machine learning is the study of letting computers learn to perform a specific task in a datadriven manner. In particular, Bayesian machine learning has attracted enor mous attention mainly due to their ability to provide uncertainty estimates following Bayesian inference. This thesis focuses on Gaussian processes (GPs), a rich class of Bayesian nonparametric models for performing Bayesian machine learning with formal measures of predictive uncertainty.
However, the applicability of GP in large datasets and in hierarchical composition of GPs is severely limited by computational issues and intractabilities. Therefore, it is crucial to develop accurate and efficient inference algorithms to address these challenges. To this end, this thesis aims at proposing a series of novel approximate Bayesian inference methods for a wide variety of GP models, which unifies the previous literatures, significantly extends them and hopefully lays the foundation for future inference methods.
To start with, this thesis presents a unifying perspective of existing inducing variablesbased GP models, sparse GP (SGP) models and variational inference for SGP models (VSGP). Then, to further mitigate the issue of overfitting during optimization, we present a novel variational inference framework for deriving a family of Bayesian SGP regression models, referred to as variational Bayesian SGP (VBSGP) regression models.
Next, taking into account the fact that the expressiveness of GP and SGP depends heavily on the design of the kernel function, we further extend the expressive power of GP by introducing Deep GP (DGP), which is a hierarchical composition of GP models. Unfortunately, exact inference in DGP is intractable, which has motivated the recent development of deterministic and stochastic approximation methods. However, the deterministic approximation methods yield a biased posterior belief while the stochastic one is computationally costly. In this regard, we present the implicit posterior variational inference (IPVI) framework for DGPs that can ideally recover an unbiased posterior belief and still preserve time efficiency. Inspired by generative adversarial networks, our IPVI framework casts the DGP inference problem as a two player game in which a Nash equilibrium, interestingly, coincides with an unbiased posterior belief.
We hope this thesis at least provides additional confidence and clarity for researchers who are devoting themselves to Bayesian nonparametric models, Gaussian process models in particular. Moreover, we also wish this thesis to offer inspirations for future works, and some thoughts that could be useful for future solutions.
 Implicit Posterior Variational Inference for Deep Gaussian Processes.
Haibin Yu, Yizhou Chen, Zhongxiang Dai, Kian Hsiang Low & Patrick Jaillet.
In Advances in Neural Information Processing Systems 32: 33rd Annual Conference on Neural Information Processing Systems (NeurIPS'19), pages 1447514486, Vancouver, Canada, Dec 712, 2019.
3% acceptance rate (spotlight presentation)
Abstract. A multilayer deep Gaussian process (DGP) model is a hierarchical composition of GP models with a greater expressive power. Exact DGP inference is intractable, which has motivated the recent development of deterministic and stochastic approximation methods. Unfortunately, the deterministic approximation methods yield a biased posterior belief while the stochastic one is computationally costly. This paper presents an implicit posterior variational inference (IPVI) framework for DGPs that can ideally recover an unbiased posterior belief and still preserve time efficiency. Inspired by generative adversarial networks, our IPVI framework achieves this by casting the DGP inference problem as a twoplayer game in which a Nash equilibrium, interestingly, coincides with an unbiased posterior belief. This consequently inspires us to devise a bestresponse dynamics algorithm to search for a Nash equilibrium (i.e., an unbiased posterior belief). Empirical evaluation shows that IPVI outperforms the stateoftheart approximation methods for DGPs.
 Stochastic Variational Inference for Bayesian Sparse Gaussian Process Regression.
Haibin Yu, Trong Nghia Hoang, Kian Hsiang Low & Patrick Jaillet.
In Proceedings of the International Joint Conference on Neural Networks (IJCNN'19), Budapest, Hungary, Jul 1419, 2019.
52.4% acceptance rate
Abstract. This paper presents a novel variational inference framework for deriving a family of Bayesian sparse Gaussian process regression (SGPR) models whose approximations are variationally optimal with respect to the fullrank GPR model enriched with various corresponding correlation structures of the observation noises. Our variational Bayesian SGPR (VBSGPR) models jointly treat both the distributions of the inducing variables and hyperparameters as variational parameters, which enables the decomposability of the variational lower bound that in turn can be exploited for stochastic optimization. Such a stochastic optimization involves iteratively following the stochastic gradient of the variational lower bound to improve its estimates of the optimal variational distributions of the inducing variables and hyperparameters (and hence the predictive distribution) of our VBSGPR models and is guaranteed to achieve asymptotic convergence to them. We show that the stochastic gradient is an unbiased estimator of the exact gradient and can be computed in constant time per iteration, hence achieving scalability to big data. We empirically evaluate the performance of our proposed framework on two realworld, massive datasets.
 A Generalized Stochastic Variational Bayesian Hyperparameter Learning Framework for Sparse Spectrum Gaussian Process Regression.
Quang Minh Hoang, Trong Nghia Hoang & Kian Hsiang Low.
In Proceedings of the 31st AAAI Conference on Artificial Intelligence (AAAI17), pages 20072014, San Francisco, CA, Feb 49, 2017.
24.6% acceptance rate (oral presentation)
Abstract. While much research effort has been dedicated to scaling up sparse Gaussian process (GP) models based on inducing variables for big data, little attention is afforded to the other less explored class of lowrank GP approximations that exploit the sparse spectral representation of a GP kernel. This paper presents such an effort to advance the state of the art of sparse spectrum GP models to achieve competitive predictive performance for massive datasets. Our generalized framework of stochastic variational Bayesian sparse spectrum GP (sVBSSGP) models addresses their shortcomings by adopting a Bayesian treatment of the spectral frequencies to avoid overfitting, modeling these frequencies jointly in its variational distribution to enable their interaction a posteriori, and exploiting local data for boosting the predictive performance. However, such structural improvements result in a variational lower bound that is intractable to be optimized. To resolve this, we exploit a variational parameterization trick to make it amenable to stochastic optimization. Interestingly, the resulting stochastic gradient has a linearly decomposable structure that can be exploited to refine our stochastic optimization method to incur constant time per iteration while preserving its property of being an unbiased estimator of the exact gradient of the variational lower bound. Empirical evaluation on realworld datasets shows that sVBSSGP outperforms stateoftheart stochastic implementations of sparse GP models.
 A Unifying Framework of Anytime Sparse Gaussian Process Regression Models with Stochastic Variational Inference for Big Data.
Trong Nghia Hoang, Quang Minh Hoang & Kian Hsiang Low.
In Proceedings of the 32nd International Conference on Machine Learning (ICML15), pages 569578, Lille, France, Jul 611, 2015.
26.0% acceptance rate
Abstract. This paper presents a novel unifying framework of anytime sparse Gaussian process regression (SGPR) models that can produce good predictive performance fast and improve their predictive performance over time. Our proposed unifying framework reverses the variational inference procedure to theoretically construct a nontrivial, concave functional that is maximized at the predictive distribution of any SGPR model of our choice.
As a result, a stochastic natural gradient ascent method can be derived that involves iteratively following the stochastic natural gradient of the functional to improve its estimate of the predictive distribution of the chosen SGPR model
and is guaranteed to achieve asymptotic convergence to it. Interestingly, we show that if the predictive distribution of the chosen SGPR model
satisfies certain decomposability conditions, then the stochastic natural gradient is an unbiased estimator of the exact natural gradient and can be computed in constant time (i.e., independent of data size) at each iteration. We empirically evaluate the tradeoff between the predictive performance vs. time efficiency of the anytime SGPR models on two realworld millionsized datasets.
 New Advances on Bayesian and DecisionTheoretic Approaches for Interactive Machine Learning.
Trong Nghia Hoang.
Ph.D. Thesis, Department of Computer Science, National University of Singapore, Feb 2015.
Abstract.
The explorationexploitation tradeoff is a fundamental dilemma in many interactive learning scenarios which include both aspects of reinforcement learning (RL) and active learning (AL): An autonomous agent, situated in an unknown environment, has to actively extract knowledge from the environment by taking actions (or conducting experiments) based on its previously collected information to make accurate predictions or to optimize some utility functions. Thus, to make the most effective use of their resourceconstrained budget (e.g., processing time, experimentation cost), the agent must choose carefully between (a) exploiting options (e.g., actions, experiments) which are recommended by its current, possibly incomplete model of the environment, and (b) exploring the other ostensibly suboptimal choices to gather more information.
For example, an RL agent has to face a dilemma between (a) exploiting the mostrewarding action according to the current statistical model of the environment at the risk of running into catastrophic situations if the model is not accurate, and (b) exploring a suboptimal action to gather more information so as to improve the model's accuracy at the potential price of losing the shortterm reward. Similarly, an AL algorithm/agent has to consider between (a) conducting the most informative experiments according to its current estimation of the environment model's parameters (i.e., exploitation), and (b) running experiments that help improving the estimation accuracy of these parameters (i.e., exploration).
More often, learning strategies that ignore exploration will likely exhibit suboptimal performance due to their imperfect knowledge while, conversely, those that entirely focus on exploration might suffer the cost of learning without benefitting from it. Therefore, a good explorationexploitation tradeoff is critical to the success of those interactive learning agents: In order to perform well, they must strike the right balance between these two conflicting objectives. Unfortunately, while this tradeoff has been wellrecognized since the early days of RL, the studies of explorationexploitation have mostly been developed for theoretical settings in the respective field of RL and, perhaps surprisingly, glossed over in the existing AL literature. From a practical point of view, we see three limiting factors:
 Previous works addressing the explorationexploitation tradeoff in RL have largely focused on simple choices of the environment model and consequently, are not practical enough to accommodate realworld applications that have far more complicated environment structures. In fact, we find that most recent advances in Bayesian reinforcement learning (BRL) have only been able to analytically trade off between exploration and exploitation under a simple choice of models such as FlatDirichletMultinomial (FDM) whose independence and modeling assumptions do not hold for many realworld applications.
 Nearly all of the notable works in the AL literature primarily advocate the use of greedy/myopic algorithms whose rates of convergence (i.e., the number of experiments required by the learning algorithm to achieve a desired performance in the worst case) are provably minimax optimal for simple classes of learning tasks (e.g., threshold learning). While these results have greatly ad vanced our understanding about the limit of myopic AL in worstcase scenarios, significantly less is presently known about whether it is possible to devise nonmyopic AL strategies which optimize the explorationexploitation tradeoff to achieve the best expected performance in budgeted learning scenarios.
 The issue of scalability of the existing predictive models (e.g., Gaussian processes) used in AL has generally been underrated since the majority of literature considers smallscale environments which only consist of a few thousand candidate experiments to be selected by singlemode AL algorithms one at a time prior to retraining the model. In contrast, largescale environments usually have a massive set of million candidate experiments among which tens or hundreds of thousands should be actively selected for learning. For such dataintensive problems, it is often more costeffective to consider batchmode AL algorithms which select and conduct multiple experiments in parallel at each stage to collect observations in batch. Retraining the predictive model after incorporating each batch of observations then becomes a computational bottleneck as the collected dataset at each stage quickly grows up to tens or even hundreds of thousand data points.
This thesis outlines some recent progresses that we have been able to make while working toward satisfactory answers to the above challenges, along with practical algorithms that achieve them:
 In particular, in order to put BRL into practice for more complicated and practical problems, we propose a novel framework called Interactive Bayesian Reinforcement Learning (IBRL) to integrate the general class of parametric models and model priors, thus allowing the practitioners' domain knowledge to be exploited to produce a finegrained and compact representation of the environment as often required in many realworld applications. Interestingly, we show how the nonmyopic Bayesoptimal policy can be derived analytically by solving IBRL exactly and propose an approximation algorithm to compute it efficiently in polynomial time. Our empirical studies show that the proposed approach performs competitively with the existing stateoftheart algorithms.
 Then, to establish a theoretical foundation for the explorationexploitation tradeoff in singlemode active learning scenarios with resourceconstrained budgets, we present a novel ϵBayesoptimal DecisionTheoretic Active Learning (ϵBAL) framework which advocates the use of differential entropy as a performance measure and consequently, derives a learning policy that can approximate the optimal expected performance arbitrarily closely (i.e., within an arbitrary loss bound ϵ). To meet the realtime requirement in timecritical applications, we then propose an asymptotically ϵoptimal, branchandbound anytime algorithm based on ϵBAL with performance guarantees. In practice, we empirically demonstrate with both synthetic and realworld datasets that the proposed approach outperforms the stateoftheart algorithms in budgeted scenarios.
 Lastly, to facilitate the future developments of largescale, nonmyopic AL applications, we further introduce a highly scalable family of anytime predictive models for AL which provably converge toward a wellknown class of sparse Gaussian processes (SGPs). Unlike the existing predictive models of AL which cannot be updated incrementally and are only capable of processing middlesized datasets (i.e., a few thousands of data points), our proposed models can process massive datasets in an anytime fashion, thus providing a principled tradeoff between the processing time and the predictive accuracy. The efficiency of our framework is then demonstrated empirically on a variety of largescale realworld datasets which contains hundreds of thousand data points.
 GPLocalize: Persistent Mobile Robot Localization using Online Sparse Gaussian Process Observation Model.
Nuo Xu, Kian Hsiang Low, Jie Chen, Keng Kiat Lim & Etkin Baris Ozgul.
In Proceedings of the 28th AAAI Conference on Artificial Intelligence (AAAI14), pages 25852592, Quebec City, Canada, Jul 2731, 2014.
16.6% acceptance rate (oral presentation)
Also appeared in
RSS14 Workshop on NonParametric Learning in Robotics, Berkeley, CA, Jul 12, 2014.
Abstract. Central to robot exploration and mapping is the task of persistent localization in environmental fields characterized by spatially correlated measurements. This paper presents a Gaussian process localization (GPLocalize) algorithm that, in contrast to existing works, can exploit the spatially correlated field measurements taken during a robot's exploration (instead of relying on prior training data) for efficiently and scalably learning the GP observation model online through our proposed novel online sparse GP. As a result, GPLocalize is capable of achieving constant time and memory (i.e., independent of the size of the data) per filtering step, which demonstrates the practical feasibility of using GPs for persistent robot localization and autonomy. Empirical evaluation via simulated experiments with realworld datasets and a real robot experiment shows that GPLocalize outperforms existing GP localization algorithms.
 Generalized Online Sparse Gaussian Processes with Application to Persistent Mobile Robot Localization.
Kian Hsiang Low, Nuo Xu, Jie Chen, Keng Kiat Lim & Etkin Baris Ozgul.
In T. Calders, F. Esposito, E. Hüllermeier, R. Meo, editors, Machine Learning and Knowledge Discovery in Databases  European Conference, ECML/PKDD14 Nectar (New Scientific and Technical Advances in Research) Track, Part III, LNCS 8726, pages 499503, Springer Berlin Heidelberg, Nancy, France, Sep 1519, 2014.
Abstract. This paper presents a novel online sparse Gaussian process (GP) approximation method that is capable of achieving constant time and memory (i.e., independent of the size of the data) per time step. We theoretically guarantee its predictive performance to be equivalent to that of a sophisticated offline sparse GP approximation method. We empirically demonstrate the practical feasibility of using our online sparse GP approximation method through a realworld persistent mobile robot localization experiment.
PARALLEL AND DISTRIBUTED SPARSE GAUSSIAN PROCESSES FOR BIG DATA
PROJECT DURATION : Aug 2010  Present
PROJECT AFFILIATION

SingaporeMIT Alliance for Research and Technology (SMART) Future Urban Mobility (FM) IRG (Collaborator: Patrick Jaillet, MIT)

SensorEnhanced Social Media (SeSaMe) Centre (Collaborator: Mohan Kankanhalli)
PROJECT FUNDING
 RIE2020 AME Programmatic Fund : Learning with Less Data, SGD $1,218,600, Apr 2021  Mar 2024
 MOE AcRF Tier 1 Reimagine Research Scheme Funding : Scalable AI Phenome Platform towards FastForward Plant Breeding (Machine Learning),
SGD $348,600, Mar 2021  Mar 2024
 RIE2020 AME IAFPP : High Performance Precision Agriculture (HiPPA) System, S$1,197,960, Mar 2020  Feb 2024
 MOE AcRF Tier 2 Grant : Scaling up Gaussian Process Predictive Models for Big Data, SGD $737,461, Jul 2017  Jul 2020
 SMART Subaward Agreements  FM IRG :
Spatiotemporal Modeling and Prediction of Traffic Patterns,
SGD $361,456.17, Oct 2011  Mar 2017
 Research Collaboration Agreement with Sumitomo Electric Industries, Ltd. : Estimation/Prediction Algorithm for Traffic Volume without Rich Installation of Detectors, JPY $3,000,000, Sep 2013  Nov 2014
PROBLEM MOTIVATION
Gaussian process (GP) models are a rich class of Bayesian nonparametric models that can perform probabilistic regression by providing Gaussian predictive distributions with formal measures of the predictive uncertainty.
Unfortunately, a GP model is handicapped by its poor scalability in the size of the data, hence limiting its practical use to small data. To improve its scalability, two families of sparse GP regression methods have been proposed: (a) Lowrank approximate representations
of the fullrank GP (FGP) model are wellsuited for modeling slowlyvarying functions with large correlation and can use all the data for predictions. But, they require a relatively high rank to capture smallscale features/patterns (i.e., of small correlation) with high fidelity, thus losing their computational advantage. (b) In contrast, localized regression and covariance tapering methods (e.g., local GPs and compactly supported covariance functions) are particularly useful for modeling rapidlyvarying functions with small correlation. However, they can only utilize local data for predictions, thereby performing poorly in input regions with little/no data. Furthermore, to accurately represent largescale features/patterns (i.e., of large correlation), the locality/tapering range has to be increased considerably, thus sacrificing their time efficiency.
Recent sparse GP regression methods have unified approaches from the two families described above to harness their complementary modeling and predictive capabilities (hence, eliminating their deficiencies) while retaining their computational advantages. Specifically, after approximating the FGP (in particular, its covariance matrix) with a lowrank representation based on the notion of inducing variables, a sparse covariance matrix approximation of the resulting residual process is made. However, this sparse residual covariance matrix approximation imposes a fairly strong conditional independence assumption given the inducing variables since the number of inducing variables cannot be too large to preserve time efficiency. We argue in this work that such a strong assumption is an overkill: It is in fact possible to construct a more refined, dense residual covariance matrix approximation by exploiting a Markov assumption and, perhaps surprisingly, still achieve scalability, which distinguishes our work here from existing sparse GP regression methods utilizing lowrank representations (i.e., including the unified approaches) described earlier.
As a result, our proposed residual covariance matrix approximation can significantly relax the conditional independence assumption (especially with larger data), hence potentially improving the predictive performance.
PROPOSED METHODOLOGY
This work presents a lowrankcumMarkov approximation (LMA) of the FGP model that is novel in leveraging the dual computational advantages stemming from complementing the reducedrank covariance matrix approximation based on the inducing variables with the residual covariance matrix approximation due to the Markov assumption;
the latter approximation is guaranteed to be closest in the KullbackLeibler distance criterion subject to some constraint. Consequently, our proposed LMA method can trade off between the number of inducing variables and the order of the Markov property to (a) incur lower computational cost than sparse GP regression methods utilizing lowrank representations with only the number of inducing variables or spectral points as the varying parameter while achieving predictive performance comparable to them and (b) accurately represent features/patterns of any scale.
Interestingly, varying the Markov order produces a spectrum of LMAs with the partially independent conditional (PIC) approximation and FGP at the two extremes. An important advantage of LMA over most existing sparse GP regression methods is that it is amenable to parallelization on multiple machines/cores, thus gaining greater scalability for performing realtime predictions necessary in many timecritical applications and decision support systems (e.g., ocean sensing, traffic monitoring). Our parallel LMA method is implemented using the message passing interface (MPI) framework to run in clusters of up to 32 computing nodes and its predictive performance, scalability, and speedup are empirically evaluated on three realworld datasets (i.e., including a millionsized dataset).
PUBLICATIONS
 A Distributed Variational Inference Framework for Unifying Parallel Sparse Gaussian Process Regression Models.
Trong Nghia Hoang, Quang Minh Hoang & Kian Hsiang Low.
In Proceedings of the 33rd International Conference on Machine Learning (ICML16), pages 382391, New York City, NY, Jun 1924, 2016.
24.3% acceptance rate
Abstract. This paper presents a novel distributed variational inference framework that unifies many parallel sparse Gaussian process regression (SGPR) models for scalable hyperparameter learning with big data. To achieve this, our framework exploits a structure of correlated noise process model that represents the observation noises as a finite realization of a highorder Gaussian Markov random process. By varying the Markov order and covariance function for the noise process model, different variational SGPR models result. This consequently allows the correlation structure of the noise process model to be characterized for which a particular variational SGPR model is optimal. We empirically evaluate the predictive performance and scalability of the distributed variational SGPR models unified by our framework on two realworld datasets.
 Parallel Gaussian Process Regression for Big Data: LowRank Representation Meets Markov Approximation.
Kian Hsiang Low, Jiangbo Yu, Jie Chen & Patrick Jaillet.
In Proceedings of the 29th AAAI Conference on Artificial Intelligence (AAAI15), pages 28212827, Austin, TX, Jan 2529, 2015.
26.67% acceptance rate
Abstract. The expressive power of a Gaussian process (GP) model comes at a cost of poor scalability in the data size.
To improve its scalability, this paper presents a lowrankcumMarkov approximation (LMA) of the GP model that is novel in leveraging the dual computational advantages stemming from complementing a lowrank approximate representation of the fullrank GP based on a support set of inputs with a Markov approximation of the resulting residual process; the latter approximation is guaranteed to be closest in the KullbackLeibler distance criterion subject to some constraint
and is considerably more refined than that of existing sparse GP models utilizing lowrank representations due to its more relaxed conditional independence assumption (especially with larger data).
As a result, our LMA method can trade off between the size of the support set and the order of the Markov property to (a) incur lower computational cost than such sparse GP models while achieving predictive performance comparable to them and (b) accurately represent features/patterns of any scale.
Interestingly, varying the Markov order produces a spectrum of LMAs
with PIC approximation and fullrank GP at the two extremes.
An advantage of our LMA method is that it is amenable to parallelization on multiple machines/cores, thereby gaining greater scalability.
Empirical evaluation on three realworld datasets in clusters of up to 32 computing nodes shows that our centralized and parallel LMA methods are significantly more timeefficient and scalable than stateoftheart sparse and fullrank GP regression methods
while achieving comparable predictive performances.
 Parallel Gaussian Process Regression with LowRank Covariance Matrix Approximations.
Jie Chen, Nannan Cao, Kian Hsiang Low, Ruofei Ouyang, Colin KengYan Tan & Patrick Jaillet.
In Proceedings of the 29th Conference on Uncertainty in Artificial Intelligence (UAI13), pages 152161, Bellevue, WA, Jul 1115, 2013.
31.3% acceptance rate
Abstract. Gaussian processes (GP) are Bayesian nonparametric models that are widely used for probabilistic regression. Unfortunately, it cannot scale well with large data nor perform realtime predictions due to its cubic time cost in the data size. This paper presents two parallel GP regression methods that exploit lowrank covariance matrix approximations for distributing the computational load among parallel machines to achieve time efficiency and scalability. We theoretically guarantee the predictive performances of our proposed parallel GPs to be equivalent to that of some centralized approximate GP regression methods: The computation of their centralized counterparts can be distributed among parallel machines, hence achieving greater time efficiency and scalability. We analytically compare the properties of our parallel GPs such as time, space, and communication complexity. Empirical evaluation on two realworld datasets in a cluster of 20 computing nodes shows that our parallel GPs are significantly more timeefficient and scalable than their centralized counterparts and exact/full GP while achieving predictive performances comparable to full GP.
 Gaussian ProcessBased Decentralized Data Fusion and Active Sensing Agents: Towards LargeScale Modeling and Prediction of Spatiotemporal Traffic Phenomena.
Jie Chen.
Ph.D. Thesis, Department of Computer Science, National University of Singapore, Dec 2013.
Abstract.
Knowing and understanding the environmental phenomena is important to many real world applications. This thesis is devoted to study largescale modeling and prediction of spatiotemporal environmental phenomena (i.e., urban traffic phenomena). Towards this goal, our proposed approaches rely on a class of Bayesian nonparametric models: Gaussian processes (GP).
To accurately model spatiotemporal urban traffic phenomena in real world situation, a novel relational GP taking into account both the road segment features and road network topology information is proposed to model real world traffic conditions over road network. Additionally, a GP variant called logGaussian process (lGP) is exploited to model an urban mobility demand pattern which contains skewness and extremity in demand measurements.
To achieve efficient and scalable urban traffic phenomenon prediction given a large phenomenon data, we propose three novel parallel GPs: parallel partially independent training conditional (pPITC), parallel partially independent conditional(pPIC) and parallel incomplete Cholesky factorization (pICF)based approximations of GP model, which can distribute their computational load into a cluster of parallel/multicore machines, thereby achieving time efficiency. The predictive performances of such parallel GPs are theoretically guaranteed to be equivalent to that of some centralized approaches to approximate full/exact GP regression. The proposed parallel GPs are implemented using the message passing interface (MPI) framework and tested on two large real world datasets. The theoretical and empirical results show that our parallel GPs achieve significantly better time efficiency and scalability than that of full GP, while achieving comparable accuracy. They also achieve fine speedup performance that is the ratio of time required by the parallel algorithms and their centralized counterparts.
To exploit active mobile sensors to perform decentralized perception of the spatiotemporal urban traffic phenomenon, we propose a decentralized algorithm framework: Gaussian processbased decentralized data fusion and active sensing (D2FAS) which is composed of a decentralized data fusion (DDF) component and a decentralized active sensing (DAS) component. The DDF component includes a novel Gaussian processbased decentralized data fusion (GPDDF) algorithm that can achieve remarkably efficient and scalable prediction of phenomenon and a novel Gaussian processbased decentralized data fusion with local augmentation (GPDDF+) algorithm that can achieve better predictive accuracy while preserving time efficiency of GPDDF. The predictive performances of both GPDDF and GPDDF+ are theoretically guaranteed to be equivalent to that of some sophisticated centralized sparse approximations of exact/full GP. For the DAS component, we propose a novel partially decentralized active sensing (PDAS) algorithm that exploits property in correlation structure of GPDDF to enable mobile sensors cooperatively gathering traffic phenomenon data along a nearoptimal joint walk with theoretical guarantee, and a fully decentralized active sensing (FDAS) algorithm that guides each mobile sensor gather phenomenon data along its locally optimal walk.
Lastly, to justify the practicality of the D2FAS framework, we develop and test D2FAS algorithms running with active mobile sensors on real world datasets for monitoring traffic conditions and sensing/servicing urban mobility demands. Theoretical and empirical results show that the proposed algorithms are significantly more timeefficient, more scalable in the size of data and in the number of sensors than the stateoftheart centralized approaches, while achieving comparable predictive accuracy.
PRESENTATIONS
 Gaussian ProcessBased Decentralized Data Fusion
and Active Sensing Agents:
Towards LargeScale Modeling & Prediction of Spatiotemporal Traffic Phenomena.
Kian Hsiang Low.
Invited speaker at the RSS13 Workshop on Robotic Exploration, Monitoring, and Information Collection: Nonparametric Modeling, Informationbased Control, and Planning under Uncertainty, Berlin, Germany, Jun 2728, 2013.
INTENTIONAWARE PLANNING UNDER UNCERTAINTY FOR INTERACTING OPTIMALLY WITH SELFINTERESTED AGENTS
PROJECT DURATION : May 2011  Present
PROJECT AFFILIATION :
SingaporeMIT Alliance for Research and Technology (SMART) Future Urban Mobility (FM) IRG (Collaborators: Emilio Frazzoli, MIT; Daniela Rus, MIT)
PROJECT FUNDING : SMART Subaward Agreements  FM IRG :
Autonomy in MobilityOnDemand Systems,
SGD $1,348,638.22, Aug 2010  Dec 2015
PROBLEM MOTIVATION
Designing and developing efficient planning algorithms for intelligent agents to interact and perform effectively among other selfinterested agents has recently emerged as a grand challenge in noncooperative multiagent systems. Such a challenge is posed by many realworld applications, which include automated electronic trading markets where software agents interact, and traffic intersections where autonomous cars have to negotiate with humandriven vehicles to cross them, among others. Modeling, predicting, and learning the other agents' intentions efficiently is therefore critical to overcoming this challenge.
In practice, it is highly nontrivial to model and predict the other agents' intentions efficiently. Existing works addressing this challenge are often undermined due to either the restrictive assumptions on the agents' behaviors or the prohibitively expensive cost of modeling and predicting their intentions:
 Gametheoretic approaches tend to assume the agents' behaviors to be perfectly rational using the wellfounded solution concepts of classical game theory such as Nash equilibrium that suffers from the following drawbacks: (a) Multiple equilibria may exist, (b) only the optimal actions corresponding to the equilibria are specified, and (c) they assume that the agents do not collaborate to beneficially deviate from the equilibrium, which is often violated by human agents.
 In contrast, decisiontheoretic approaches propose to extend singleagent sequential decision making frameworks under partial observability such as POMDP to explicitly characterize the bounded rationality of selfinterested agents. In particular, Interactive POMDP (IPOMDP) replaces a POMDP's flat belief over physical states with an interactive belief of k levels of hierarchy over both the physical state space and the other agent's beliefs, the latter of which are recursively defined as interactive beliefs of k1 levels of hierarchy.
As a consequence, the agent's "optimal" behavior computed at hierarchical level k is expected to be the best response to the other agent's expected behavior at hierarchical level k1. This surprisingly coincides with the wellfounded cognitive hierarchy model of games where k is referred to as the reasoning depth. Here, the bounded rationality of the agents is explicitly accounted for by making k finite and defining their expected behaviors at level 0 as uniformly random. Empowered by such enriched and highly expressive belief space, IPOMDP can explicitly model and predict the other agent's intention. Unfortunately, solving IPOMDP (e.g., solving for the agent's expected behavior at level k) is fraught with computational curses of dimensionality, history, and nested reasoning due to its highly sophisticated structure.
Furthermore, there is another practical concern for these decisiontheoretic approaches: They often require the behavioral model's parameters (e.g., hierarchical level k) to be completely specified by the practitioners, which can be very impractical in many situations where it is nontrivial to do so or the prior knowledge is insufficient to reliably derive these parameters. This essentially boils down to the need of learning while interacting with the other selfinterested agents and, interestingly, an exploitationexploration tradeoff to be made while doing so: Should an agent exploit the "best" action based on its (possibly misleading) knowledge to maximize the payoff or explore a "suboptimal" action to refine its knowledge?
Naively, one may attempt to solve it by directly tapping into the huge body of existing works in Bayesian Reinforcement Learning (BRL), which offers a broad range of principled treatments of this issue under singleagent contexts. However, most of these works often assume very simple and specific parameterizations of the unknown environments, thus rendering them inapplicable to the context where the other agent's behavior has a far more complicated parameterization. More importantly, the other agent's behavior often needs to be modeled differently depending on the specific context. Grounding in the context of existing BRL frameworks, either the domain expert struggles to best fit his prior knowledge to the supported set of parameterizations or the agent developer has to redesign the framework to incorporate a new modeling scheme. Arguably, there is no free lunch when it comes to modeling the agent's behavior across various applications.
The main focus of our work here is thus to investigate and address the following questions:
 How can intention prediction be efficiently exploited and made practical in planning under partial observability? In particular, how can the bounded rationality of the other agents be explicitly modeled without incurring prohibitive computational cost?
 How can existing BRL frameworks be refined to allow a domain expert to freely incorporate his choice of design in modeling the other agents' behaviors?
This question is signficant in putting theory into practice and, when answered, can potentially bridge the gap between learning in single and (selfinterested) multiagent systems.
PROPOSED METHODOLOGY
 To answer the first question, we first develop a novel intentionaware nested MDP framework for planning in fully observable multiagent environments. Inspired by the cognitive hierarchy model of games, nested MDP constitutes a recursive reasoning formalism to predict the other agent's intention and then exploit it to plan our agent's optimal interaction policy. We show that nested MDP incurs linear time in the planning horizon length and reasoning depth. Then, we propose an intentionaware IPOMDP Lite framework for planning in partially observable multiagent environments that, in particular, exploits a practical structural assumption: The intention of the other agent is driven by nested MDP, which is demonstrated theoretically to be an effective surrogate of its true intention when the agents have fine sensing and actuation capabilities. This assumption will allow the other agent's intention to be predicted efficiently and, consequently, IPOMDP Lite to be solved effectively, as demonstrated theoretically and empirically in our work.
 To tackle the second question, we present a novel generalization of BRL called Interactive BRL (IBRL) to integrate any parametric model and model prior of the other agent's behavior specified by domain experts, thus effectively allowing the other agent's sophisticated behavior to be represented in a finegrained manner based on the practitioners' prior knowledge. In particular, we show how the nonmyopic Bayesoptimal policy can be derived analytically by solving IBRL exactly and propose an approximation algorithm to compute it efficiently in polynomial time. Empirically, we demonstrate IBRL's performance using an interesting traffic problem modeled after a realworld situation.
PUBLICATIONS
 New Advances on Bayesian and DecisionTheoretic Approaches for Interactive Machine Learning.
Trong Nghia Hoang.
Ph.D. Thesis, Department of Computer Science, National University of Singapore, Feb 2015.
Abstract.
The explorationexploitation tradeoff is a fundamental dilemma in many interactive learning scenarios which include both aspects of reinforcement learning (RL) and active learning (AL): An autonomous agent, situated in an unknown environment, has to actively extract knowledge from the environment by taking actions (or conducting experiments) based on its previously collected information to make accurate predictions or to optimize some utility functions. Thus, to make the most effective use of their resourceconstrained budget (e.g., processing time, experimentation cost), the agent must choose carefully between (a) exploiting options (e.g., actions, experiments) which are recommended by its current, possibly incomplete model of the environment, and (b) exploring the other ostensibly suboptimal choices to gather more information.
For example, an RL agent has to face a dilemma between (a) exploiting the mostrewarding action according to the current statistical model of the environment at the risk of running into catastrophic situations if the model is not accurate, and (b) exploring a suboptimal action to gather more information so as to improve the model's accuracy at the potential price of losing the shortterm reward. Similarly, an AL algorithm/agent has to consider between (a) conducting the most informative experiments according to its current estimation of the environment model's parameters (i.e., exploitation), and (b) running experiments that help improving the estimation accuracy of these parameters (i.e., exploration).
More often, learning strategies that ignore exploration will likely exhibit suboptimal performance due to their imperfect knowledge while, conversely, those that entirely focus on exploration might suffer the cost of learning without benefitting from it. Therefore, a good explorationexploitation tradeoff is critical to the success of those interactive learning agents: In order to perform well, they must strike the right balance between these two conflicting objectives. Unfortunately, while this tradeoff has been wellrecognized since the early days of RL, the studies of explorationexploitation have mostly been developed for theoretical settings in the respective field of RL and, perhaps surprisingly, glossed over in the existing AL literature. From a practical point of view, we see three limiting factors:
 Previous works addressing the explorationexploitation tradeoff in RL have largely focused on simple choices of the environment model and consequently, are not practical enough to accommodate realworld applications that have far more complicated environment structures. In fact, we find that most recent advances in Bayesian reinforcement learning (BRL) have only been able to analytically trade off between exploration and exploitation under a simple choice of models such as FlatDirichletMultinomial (FDM) whose independence and modeling assumptions do not hold for many realworld applications.
 Nearly all of the notable works in the AL literature primarily advocate the use of greedy/myopic algorithms whose rates of convergence (i.e., the number of experiments required by the learning algorithm to achieve a desired performance in the worst case) are provably minimax optimal for simple classes of learning tasks (e.g., threshold learning). While these results have greatly advanced our understanding about the limit of myopic AL in worstcase scenarios, significantly less is presently known about whether it is possible to devise nonmyopic AL strategies which optimize the explorationexploitation tradeoff to achieve the best expected performance in budgeted learning scenarios.
 The issue of scalability of the existing predictive models (e.g., Gaussian processes) used in AL has generally been underrated since the majority of literature considers smallscale environments which only consist of a few thousand candidate experiments to be selected by singlemode AL algorithms one at a time prior to retraining the model. In contrast, largescale environments usually have a massive set of million candidate experiments among which tens or hundreds of thousands should be actively selected for learning. For such dataintensive problems, it is often more costeffective to consider batchmode AL algorithms which select and conduct multiple experiments in parallel at each stage to collect observations in batch. Retraining the predictive model after incorporating each batch of observations then becomes a computational bottleneck as the collected dataset at each stage quickly grows up to tens or even hundreds of thousand data points.
This thesis outlines some recent progresses that we have been able to make while working toward satisfactory answers to the above challenges, along with practical algorithms that achieve them:
 In particular, in order to put BRL into practice for more complicated and practical problems, we propose a novel framework called Interactive Bayesian Reinforcement Learning (IBRL) to integrate the general class of parametric models and model priors, thus allowing the practitioners' domain knowledge to be exploited to produce a finegrained and compact representation of the environment as often required in many realworld applications. Interestingly, we show how the nonmyopic Bayesoptimal policy can be derived analytically by solving IBRL exactly and propose an approximation algorithm to compute it efficiently in polynomial time. Our empirical studies show that the proposed approach performs competitively with the existing stateoftheart algorithms.
 Then, to establish a theoretical foundation for the explorationexploitation tradeoff in singlemode active learning scenarios with resourceconstrained budgets, we present a novel ϵBayesoptimal DecisionTheoretic Active Learning (ϵBAL) framework which advocates the use of differential entropy as a performance measure and consequently, derives a learning policy that can approximate the optimal expected performance arbitrarily closely (i.e., within an arbitrary loss bound ϵ). To meet the realtime requirement in timecritical applications, we then propose an asymptotically ϵoptimal, branchandbound anytime algorithm based on ϵBAL with performance guarantees. In practice, we empirically demonstrate with both synthetic and realworld datasets that the proposed approach outperforms the stateoftheart algorithms in budgeted scenarios.
 Lastly, to facilitate the future developments of largescale, nonmyopic AL applications, we further introduce a highly scalable family of anytime predictive models for AL which provably converge toward a wellknown class of sparse Gaussian processes (SGPs). Unlike the existing predictive models of AL which cannot be updated incrementally and are only capable of processing middlesized datasets (i.e., a few thousands of data points), our proposed models can process massive datasets in an anytime fashion, thus providing a principled tradeoff between the processing time and the predictive accuracy. The efficiency of our framework is then demonstrated empirically on a variety of largescale realworld datasets which contains hundreds of thousand data points.
 Interactive POMDP Lite: Towards Practical Planning to Predict and Exploit Intentions for Interacting with SelfInterested Agents.
Trong Nghia Hoang & Kian Hsiang Low.
In Proceedings of the 23rd International Joint Conference on Artificial Intelligence (IJCAI13), pages 22982305, Beijing, China, Aug 39, 2013.
13.2% acceptance rate (oral presentation)
Abstract. A key challenge in noncooperative multiagent systems is that of developing efficient planning algorithms for intelligent agents to interact and perform effectively among boundedly rational, selfinterested agents (e.g., humans). The practicality of existing works addressing this challenge is being undermined due to either the restrictive assumptions of the other agents' behavior, the failure in accounting for their rationality, or the prohibitively expensive cost of modeling and predicting their intentions. To boost the practicality of research in this field, we investigate how intention prediction can be efficiently exploited and made practical in planning, thereby leading to efficient intentionaware planning frameworks capable of predicting the intentions of other agents and acting optimally with respect to their predicted intentions. We show that the performance losses incurred by the resulting planning policies are linearly bounded by the error of intention prediction. Empirical evaluations through a series of stochastic games demonstrate that our policies can achieve better and more robust performance than the stateoftheart algorithms.
 A General Framework for Interacting BayesOptimally with SelfInterested Agents using Arbitrary Parametric Model and Model Prior.
Trong Nghia Hoang & Kian Hsiang Low.
In Proceedings of the 23rd International Joint Conference on Artificial Intelligence (IJCAI13), pages 13941400, Beijing, China, Aug 39, 2013.
28.0% acceptance rate
Abstract. Recent advances in Bayesian reinforcement learning (BRL) have shown that Bayesoptimality is theoretically achievable by modeling the environment's latent dynamics using FlatDirichletMultinomial (FDM) prior. In selfinterested multiagent environments, the transition dynamics are mainly controlled by the other agent's stochastic behavior for which FDM's independence and modeling assumptions do not hold. As a result, FDM does not allow the other agent's behavior to be generalized across different states nor specified using prior domain knowledge. To overcome these practical limitations of FDM, we propose a generalization of BRL to integrate the general class of parametric models and model priors, thus allowing practitioners' domain knowledge to be exploited to produce a finegrained and compact representation of the other agent's behavior. Empirical evaluation shows that our approach outperforms existing multiagent reinforcement learning algorithms.
 IntentionAware Planning under Uncertainty for Interacting with SelfInterested, Boundedly Rational Agents.
Trong Nghia Hoang & Kian Hsiang Low.
In Proceedings of the
11th International Conference on Autonomous Agents and MultiAgent Systems (AAMAS12), pages 12331234, Valencia, Spain, June 48, 2012.
Abstract. A key challenge in noncooperative multiagent systems is that of developing efficient planning algorithms for intelligent agents to perform effectively among boundedly rational, selfinterested (i.e., noncooperative) agents (e.g., humans). To address this challenge, we investigate how intention prediction can be efficiently exploited and made practical in planning, thereby leading to efficient intentionaware planning frameworks capable of predicting the intentions of other agents and acting optimally with respect to their predicted intentions.
GAUSSIAN PROCESS DECENTRALIZED DATA FUSION & ACTIVE SENSING AGENTS FOR MOBILITYONDEMAND SYSTEMSTowards LargeScale Spatiotemporal Traffic Modeling and Prediction
PROJECT DURATION : Aug 2010  May 2021
PROJECT AFFILIATION :
SingaporeMIT Alliance for Research and Technology (SMART) Future Urban Mobility (FM) IRG (Collaborator: Patrick Jaillet, MIT)
PROJECT FUNDING
 SMART Subaward Agreements  FM IRG :
Spatiotemporal Modeling and Prediction of Traffic Patterns,
SGD $361,456.17, Oct 2011  Mar 2017
 Research Collaboration Agreement with Sumitomo Electric Industries, Ltd. : Estimation/Prediction Algorithm for Traffic Volume without Rich Installation of Detectors, JPY $3,000,000, Sep 2013  Nov 2014
PROBLEM MOTIVATION
PRIVATE automobiles are becoming unsustainable personal mobility solutions in densely populated urban cities because the addition of parking and road spaces cannot keep pace with their escalating numbers due to limited urban land. For example, Hong Kong and Singapore have, respectively, experienced 27.6% and 37% increase in private vehicles from 2003 to 2011. However, their road networks have only expanded less than 10% in size. Without implementing sustainable measures, traffic congestions and delays will grow more severe and frequent, especially during peak hours.
Mobilityondemand (MoD) systems (e.g., Velib system of over 20000 shared bicycles in Paris, experimental carsharing systems) have recently emerged as a promising paradigm of oneway vehicle sharing for sustainable personal urban mobility, specifically, to tackle the problems of low vehicle utilization rate and parking space caused by private automobiles. Conventionally, a MoD system provides stacks and racks of light electric vehicles distributed throughout a city: When a user wants to go somewhere, he simply walks to the nearest rack, swipes a card to pick up a vehicle, drives it to the rack nearest to his destination, and drops it off. In this work, we assume the capability of a MoD system to be enhanced by deploying robotic shared vehicles (e.g., General Motors Chevrolet ENV 2.0 prototype) that can autonomously drive and cruise the streets of a densely populated urban city to be hailed by users (like taxis) instead of just waiting at the racks to be picked up. Compared to the conventional MoD system, the fleet of autonomous robotic vehicles provides greater accessibility to users who can be picked up and dropped off at any location in the road network. As a result, it can service regions of high mobility demand but with poor coverage of stacks and racks due to limited space for their installation.
The key factors in the success of a MoD system are the costs to the users and system latencies, which can be minimized by managing the MoD system effectively. To achieve this, two main technical challenges need to be addressed: (a) Realtime, finegrained mobility demand sensing and prediction, and (b) realtime active fleet management to balance vehicle supply and demand and satisfy latency requirements at sustainable operating costs. Existing works on load balancing in MoD systems, dynamic traffic assignment problems, dynamic onetoone pickup and delivery problems, and location recommendation and dispatch for cruising taxis have tackled variants of the second challenge by assuming the necessary inputs of mobility demand and traffic flow information to be perfectly or accurately known using prior knowledge or offline processing of historic data. Such an assumption does not hold for densely populated urban cities because their mobility demand patterns and traffic flow are often subject to shortterm random fluctuations and perturbations due to frequent special events (e.g., storewide sales, exhibitions), unpredictable weather conditions, unforeseen emergencies (e.g., breakdowns in public transport services), or traffic incidents (e.g., accidents, vehicle breakdowns, roadworks). So, in order for the active fleet management strategies to perform well in fleet rebalancing and route planning to service the mobility demands, they require accurate, finegrained predictive information of the spatiotemporally varying mobility demand patterns and traffic flow in real time, the former of which is the desired outcome of addressing the first challenge. To the best of our knowledge, there is little progress in the algorithm design and development to take on the first challenge, which will be a focus of our work here.
In practice, it is nontrivial to achieve realtime, accurate prediction of spatiotemporally varying traffic phenomena such as mobility demand patterns and traffic flow because the quantity of sensors that can be deployed to observe an entire road network is costconstrained. For example, static sensors such as loop detectors are traditionally placed at designated locations in a road network to collect data for predicting the traffic flow. However, they provide sparse coverage (i.e., many road segments are not observed, thus leading to data sparsity), incur high installation and maintenance costs, and cannot reposition by themselves in response to changes in the traffic flow. Lowcost GPS technology allows the collection of traffic flow data using passive mobile probes (e.g., taxis/cabs). Unlike static sensors, they can directly measure the travel times along road segments. But, they provide fairly sparse coverage due to low GPS sampling frequency (i.e., often imposed by taxi/cab companies) and no control over their routes, incur high initial implementation cost, pose privacy issues, and produce highlyvarying speeds and travel times while traversing the same road segment due to inconsistent driving behaviors. A critical mass of probes is needed on each road segment to ease the severity of the last drawback but is often hard to achieve on nonhighway segments due to sparse coverage. In contrast, we propose using the autonomous robotic vehicles as active mobile probes to overcome the limitations of static and passive mobile probes. In particular, they can be directed to explore any segments of a road network to gather realtime mobility demand data (e.g., pickup counts of different regions) and traffic flow data (e.g., speeds and travel times along road segments) at a desired GPS sampling rate while enforcing consistent driving behavior.
How then can the vacant autonomous robotic vehicles in a MoD system actively cruise a road network to gather and assimilate the most informative data for predicting a spatiotemporally varying traffic phenomenon like a mobility demand pattern or traffic flow? To solve this problem, a centralized approach to data fusion and active sensing is poorly suited because it suffers from a single point of failure and incurs huge communication, space, and time overheads with large data and fleet.
PROPOSED METHODOLOGY
Our work proposes novel decentralized data fusion and active sensing algorithms for realtime, finegrained traffic sensing, modeling, and prediction with a fleet of autonomous robotic vehicles in a MoD system. Note that the decentralized data fusion component of our proposed algorithms can also be used for static sensors and passive mobile probes.
The specific contributions of our work here include:
 Modeling and predicting a mobility demand pattern and traffic flow using, respectively, rich classes of Bayesian nonparametric models called a logGaussian process (lGP) model and a relational Gaussian process model, the latter of whose spatiotemporal correlation structure can exploit both the road segment features and road network topology information;
 Developing novel Gaussian process decentralized data fusion algorithms for cooperative perception of traffic phenomena called GPDDF and GPDDF+ whose predictive performance are theoretically guaranteed to be equivalent to that of sophisticated centralized sparse approximations of the fullrank Gaussian process (full GP) model: The computation of such sparse approximate GP models can thus be distributed among the MoD vehicles, thereby achieving efficient and scalable probabilistic prediction;
 Deriving consensus filtering variants of GPDDF and GPDDF+ that require only local communication between neighboring MoD vehicles instead of assuming alltoall communication between MoD vehicles;
 Devising decentralized active sensing algorithms (a) whose performance, when coupled with GPDDF, can be theoretically guaranteed to realize the effect of the spatiotemporal correlation structure of the traffic phenomenon and various parameter settings of the MoD system, and (b) that, when used for sampling a mobility demand pattern, can be analytically shown to exhibit, interestingly, a cruising behavior of simultaneously exploring demand hotspots and sparsely sampled regions that have higher likelihood of picking up users, hence achieving a dual effect of fleet rebalancing to service the mobility demands;
 Analyzing the time and communication overheads of our proposed algorithms: We prove that our algorithms can scale better than existing stateoftheart centralized algorithms in the size of the data and fleet;
 Empirically evaluating the predictive accuracy, time efficiency, scalability, and performance of servicing mobility demands (i.e., average cruising length of vehicles, average waiting time of users, total number of pickups) of our proposed algorithms on two datasets featuring realworld traffic phenomena such as a mobility demand pattern over the central business district of Singapore and speeds of road segments over an urban road network in Singapore.
PUBLICATIONS
 Collective Online Learning of Gaussian Processes in Massive MultiAgent Systems.
Trong Nghia Hoang, Quang Minh Hoang, Kian Hsiang Low & Jonathan P. How.
In Proceedings of the 33rd AAAI Conference on Artificial Intelligence (AAAI19), pages 78507857, Honolulu, Hawaii, Jan 27Feb 1, 2019.
16.2% acceptance rate
Abstract. This paper presents a novel Collective Online Learning of Gaussian Processes (COOLGP) framework for enabling a massive number of GP inference agents to simultaneously perform (a) efficient online updates of their GP models using their local streaming data with varying correlation structures and (b) decentralized fusion of their resulting online GP models with different learned hyperparameter settings and inducing inputs. To realize this, we exploit the notion of a common encoding structure to encapsulate the local streaming data gathered by any GP inference agent into summary statistics based on our proposed representation, which is amenable to both an efficient online update via an importance sampling trick as well as multiagent model fusion via decentralized message passing that can exploit sparse connectivity among agents for improving efficiency and enhance the robustness of our framework against transmission loss. We provide a rigorous theoretical analysis of the approximation loss arising from our proposed representation to achieve efficient online updates and model fusion. Empirical evaluations show that COOLGP is highly effective in model fusion, resilient to information disparity between agents, robust to transmission loss, and can scale to thousands of agents.
 Gaussian Process Decentralized Data Fusion Meets Transfer Learning in LargeScale Distributed Cooperative Perception.
Ruofei Ouyang & Kian Hsiang Low.
Autonomous Robots (Special Issue on MultiRobot and MultiAgent Systems), volume 44, issue 3, pages 359376, Mar 2020.
Extended version of our
AAAI18 paper
Abstract. This paper presents novel Gaussian process decentralized data fusion algorithms exploiting the notion of agentcentric support sets for distributed cooperative perception of largescale environmental phenomena. To overcome the limitations of scale in existing works, our proposed algorithms allow every mobile sensing agent to utilize a different support set and dynamically switch to another during execution for encapsulating its own data into a local summary that, perhaps surprisingly, can still be assimilated with the other agents' local summaries (i.e., based on their current choices of support sets) into a globally consistent summary to be used for predicting the phenomenon. To achieve this, we propose a novel transfer learning mechanism for a team of agents capable of sharing and transferring information encapsulated in a summary based on a support set to that utilizing a different support set with some loss that can be theoretically bounded and analyzed. To alleviate the issue of information loss accumulating over multiple instances of transfer learning, we propose a new information sharing mechanism to be incorporated into our algorithms in order to achieve memoryefficient lazy transfer learning. Empirical evaluation on three realworld datasets for up to 128 agents show that our algorithms outperform the stateoftheart methods.
 Gaussian Process Decentralized Data Fusion Meets Transfer Learning in LargeScale Distributed Cooperative Perception.
Ruofei Ouyang & Kian Hsiang Low.
In Proceedings of the 32nd AAAI Conference on Artificial Intelligence (AAAI18), pages 38763883, New Orleans, LA, Feb 28, 2018.
24.55% acceptance rate
Abstract. This paper presents novel Gaussian process decentralized data fusion algorithms exploiting the notion of agentcentric support sets for distributed cooperative perception of largescale environmental phenomena. To overcome the limitations of scale in existing works, our proposed algorithms allow every mobile sensing agent to choose a different support set and dynamically switch to another during execution for encapsulating its own data into a local summary that, perhaps surprisingly, can still be assimilated with the other agents' local summaries (i.e., based on their current choices of support sets) into a globally consistent summary to be used for predicting the phenomenon. To achieve this, we propose a novel transfer learning mechanism for a team of agents capable of sharing and transferring information encapsulated in a summary based on a support set to that utilizing a different support set with some loss that can be theoretically bounded and analyzed. To alleviate the issue of information loss accumulating over multiple instances of transfer learning, we propose a new information sharing mechanism to be incorporated into our algorithms in order to achieve memoryefficient lazy transfer learning. Empirical evaluation on realworld datasets show that our algorithms outperform the stateoftheart methods.
 Exploiting Decentralized MultiAgent Coordination for LargeScale Machine Learning Problems.
Ruofei Ouyang.
Ph.D. Thesis, Department of Computer Science, National University of Singapore, Dec 2016.
Abstract.
Nowadays, the scale of machine learning problems becomes much larger than before. It raises a huge demand in distributed perception and distributed computation. A multiagent system provides exceptional scalability for problems like active sensing and data fusion. However, many rich characteristics of largescale machine learning problems have not been addressed yet such as large input domain, nonstationarity, and high dimensionality. This thesis identifies the challenges related to these characteristics from multiagent perspective. By exploiting the correlation structure of data in largescale problems, we propose multiagent coordination schemes that can improve the scalability of the machine learning models while preserving the computation accuracy. To elaborate, the machine learning problems we are solving with multiagent coordination techniques are:
 Gaussian process regression. To perform distributed regression on a largescale environmental phenomenon, data compression is often required due to the communication costs. Currently, decentralized data fusion methods encapsulate the data into local summaries based on a fixed support set. However in a largescale field, this fixed support set, acting as a centralized component in the decentralized system, cannot approximate the correlation structure of the entire phenomenon well. It leads to evident losses in data summarization. Consequently, the regression performance will be significantly reduced.
In order to approximate the correlation structure accurately, we propose an agentcentric support set to allow every agent in the data fusion system to choose a possibly different support set and dynamically switch to another one during execution for encapsulating its own data into a local summary which, perhaps surprisingly, can still be assimilated with the other agents’ local summaries into a globally consistent summary. Together with an information sharing mechanism we designed, the new decentralized data fusion methods with agentcentric support set can be applied to regression problems on a much larger environmental phenomenon with high performance.
 Active learning. In the context of environmental sensing, active learning/active sensing is a process of taking observations to minimize the uncertainty in an environmental field. The uncertainty is quantified based on the correlation structure of the phenomenon which is traditionally assumed to be stationary for computational sake. In a largescale environmental field, this stationary assumption is often violated. Therefore, existing active sensing algorithms perform suboptimally for a nonstationary environmental phenomenon.
To the best of our knowledge, our decentralized multirobot active sensing (DECMAS) algorithm is the first work to address nonstationarity issue in the context of active sensing. The uncertainty in the phenomenon is quantified based on the nonstationary correlation structure estimated by Dirichlet process mixture of Gaussian processes. Further, our DECMAS algorithm can efficiently coordinate the exploration of multiple robots to automatically tradeoff between learning the unknown, nonstationary correlation structure and minimizing the uncertainty of the environmental phenomenon. It enables multiagent active sensing techniques to be applied to a largescale nonstationary environmental phenomenon.
 Bayesian optimization. Optimizing an unknown objective function is challenging for traditional optimization methods. Alternatively, in this situation, people use Bayesian optimization which is a modern optimization technique that can optimize a function by only utilizing the observation information (input and output values) collected through simulations. When the input dimension of the function is low, a few simulated observations can generate good result already. However, for high dimensional function, a huge number of observations are required which is impractical when the simulation consumes lots of time and resources.
Fortunately, many high dimensional problems have sparse correlation structure. Our ANOVADCOP work can decompose the correlation structure in the original highdimensional problem into many correlation structures of subsets of dimensions based on ANOVA kernel function. It significantly reduces the size of input space into a collection of lowerdimensional subspaces. Additionally, we reformulate the Bayesian optimization problem as a decentralized constrained optimization problem (DCOP) that can be efficiently solved by multiagent coordination techniques so that it can scale up to problems with hundreds of dimensions.
 Gaussian Process Decentralized Data Fusion and Active Sensing for Spatiotemporal Traffic Modeling and Prediction in MobilityonDemand Systems.
Jie Chen, Kian Hsiang Low, Patrick Jaillet & Yujian Yao.
IEEE Transactions on Automation Science and Engineering
(Special Issue on Networked Cooperative Autonomous Systems), volume 12, issue 3, pages 901921, Jul 2015.
Extended version of our UAI12 and
RSS13 papers
Abstract. Mobilityondemand (MoD) systems have recently emerged as a
promising paradigm of oneway vehicle sharing for sustainable personal
urban mobility in densely populated cities. We assume the capability of
a MoD system to be enhanced by deploying robotic shared vehicles that
can autonomously cruise the streets to be hailed by users. A key
challenge of the MoD system is that of realtime, finegrained mobility
demand and traffic flow sensing and prediction. This paper presents
novel Gaussian process (GP) decentralized data fusion and active
sensing algorithms for realtime, finegrained traffic modeling and
prediction with a fleet of MoD vehicles. The predictive performance of
our decentralized data fusion algorithms are theoretically guaranteed to
be equivalent to that of sophisticated centralized sparse GP
approximations. We derive consensus filtering variants requiring only
local communication between neighboring vehicles. We theoretically
guarantee the performance of our decentralized active sensing
algorithms. When they are used to gather informative data for mobility
demand prediction, they can achieve a dual effect of fleet rebalancing
to service mobility demands. Empirical evaluation on realworld datasets
shows that our algorithms are significantly more timeefficient and
scalable in the size of data and fleet while achieving predictive
performance comparable to that of stateoftheart algorithms.
 Gaussian ProcessBased Decentralized Data Fusion and Active Sensing Agents: Towards LargeScale Modeling and Prediction of Spatiotemporal Traffic Phenomena.
Jie Chen.
Ph.D. Thesis, Department of Computer Science, National University of Singapore, Dec 2013.
Abstract.
Knowing and understanding the environmental phenomena is important to many real world applications. This thesis is devoted to study largescale modeling and prediction of spatiotemporal environmental phenomena (i.e., urban traffic phenomena). Towards this goal, our proposed approaches rely on a class of Bayesian nonparametric models: Gaussian processes (GP).
To accurately model spatiotemporal urban traffic phenomena in real world situation, a novel relational GP taking into account both the road segment features and road network topology information is proposed to model real world traffic conditions over road network. Additionally, a GP variant called logGaussian process (lGP) is exploited to model an urban mobility demand pattern which contains skewness and extremity in demand measurements.
To achieve efficient and scalable urban traffic phenomenon prediction given a large phenomenon data, we propose three novel parallel GPs: parallel partially independent training conditional (pPITC), parallel partially independent conditional(pPIC) and parallel incomplete Cholesky factorization (pICF)based approximations of GP model, which can distribute their computational load into a cluster of parallel/multicore machines, thereby achieving time efficiency. The predictive performances of such parallel GPs are theoretically guaranteed to be equivalent to that of some centralized approaches to approximate full/exact GP regression. The proposed parallel GPs are implemented using the message passing interface (MPI) framework and tested on two large real world datasets. The theoretical and empirical results show that our parallel GPs achieve significantly better time efficiency and scalability than that of full GP, while achieving comparable accuracy. They also achieve fine speedup performance that is the ratio of time required by the parallel algorithms and their centralized counterparts.
To exploit active mobile sensors to perform decentralized perception of the spatiotemporal urban traffic phenomenon, we propose a decentralized algorithm framework: Gaussian processbased decentralized data fusion and active sensing (D2FAS) which is composed of a decentralized data fusion (DDF) component and a decentralized active sensing (DAS) component. The DDF component includes a novel Gaussian processbased decentralized data fusion (GPDDF) algorithm that can achieve remarkably efficient and scalable prediction of phenomenon and a novel Gaussian processbased decentralized data fusion with local augmentation (GPDDF+) algorithm that can achieve better predictive accuracy while preserving time efficiency of GPDDF. The predictive performances of both GPDDF and GPDDF+ are theoretically guaranteed to be equivalent to that of some sophisticated centralized sparse approximations of exact/full GP. For the DAS component, we propose a novel partially decentralized active sensing (PDAS) algorithm that exploits property in correlation structure of GPDDF to enable mobile sensors cooperatively gathering traffic phenomenon data along a nearoptimal joint walk with theoretical guarantee, and a fully decentralized active sensing (FDAS) algorithm that guides each mobile sensor gather phenomenon data along its locally optimal walk.
Lastly, to justify the practicality of the D2FAS framework, we develop and test D2FAS algorithms running with active mobile sensors on real world datasets for monitoring traffic conditions and sensing/servicing urban mobility demands. Theoretical and empirical results show that the proposed algorithms are significantly more timeefficient, more scalable in the size of data and in the number of sensors than the stateoftheart centralized approaches, while achieving comparable predictive accuracy.
 Gaussian ProcessBased Decentralized Data Fusion and Active Sensing for MobilityonDemand System.
Jie Chen, Kian Hsiang Low & Colin KengYan Tan.
In Proceedings of the
Robotics: Science and Systems Conference (RSS13), Berlin, Germany, Jun 2428, 2013.
30.1% acceptance rate
Abstract. Mobilityondemand (MoD) systems have recently emerged as a promising paradigm of oneway vehicle sharing for sustainable personal urban mobility in densely populated cities. In this paper, we enhance the capability of a MoD system by deploying robotic shared vehicles that can autonomously cruise the streets to be hailed by users. A key challenge to managing the MoD system effectively is that of realtime, finegrained mobility demand sensing and prediction. This paper presents a novel decentralized data fusion and active sensing algorithm for realtime, finegrained mobility demand sensing and prediction with a fleet of autonomous robotic vehicles in a MoD system. Our Gaussian process (GP)based decentralized data fusion algorithm can achieve a fine balance between predictive power and time efficiency. We theoretically guarantee its predictive performance to be equivalent to that of a sophisticated centralized sparse approximation for the GP model: The computation of such a sparse approximate GP model can thus be distributed among the MoD vehicles, hence achieving efficient and scalable demand prediction. Though our decentralized active sensing strategy is devised to gather the most informative demand data for demand prediction, it can achieve a dual effect of fleet rebalancing to service the mobility demands. Empirical evaluation on realworld mobility demand data shows that our proposed algorithm can achieve a better balance between predictive accuracy and time efficiency than stateoftheart algorithms.
 Decentralized Data Fusion and Active Sensing with Mobile Sensors for Modeling and Predicting Spatiotemporal Traffic Phenomena.
Jie Chen, Kian Hsiang Low, Colin KengYan Tan, Ali Oran, Patrick Jaillet, John M. Dolan & Gaurav S. Sukhatme.
In Proceedings of the 28th Conference on Uncertainty in Artificial Intelligence (UAI12), pages 163173, Catalina Island, CA, Aug 1517, 2012.
31.6% acceptance rate
Also appeared in AAMAS12 Workshop on Agents in Traffic and Transportation (ATT12), Valencia, Spain, June 48, 2012.
Abstract. The problem of modeling and predicting spatiotemporal traffic phenomena over an urban road network is important to many traffic applications such as detecting and forecasting congestion hotspots. This paper presents a decentralized data fusion and active sensing (D2FAS) algorithm for mobile sensors to actively explore the road network to gather and assimilate the most informative data for predicting the traffic phenomenon. We analyze the time and communication complexity of D2FAS and demonstrate that it can scale well with a large number of observations and sensors. We provide a theoretical guarantee on its predictive performance to be equivalent to that of a sophisticated centralized sparse approximation for the Gaussian process (GP) model: The computation of such a sparse approximate GP model can thus be parallelized and distributed among the mobile sensors (in a Googlelike MapReduce paradigm), thereby achieving efficient and scalable prediction. We also theoretically guarantee its active sensing performance that improves under various practical environmental conditions. Empirical evaluation on realworld urban road network data shows that our D2FAS algorithm is significantly more timeefficient and scalable than stateoftheart centralized algorithms while achieving comparable predictive performance.
PRESENTATIONS
 Gaussian ProcessBased Decentralized Data Fusion
and Active Sensing Agents:
Towards LargeScale Modeling & Prediction of Spatiotemporal Traffic Phenomena.
Kian Hsiang Low.
Invited speaker at the RSS13 Workshop on Robotic Exploration, Monitoring, and Information Collection: Nonparametric Modeling, Informationbased Control, and Planning under Uncertainty, Berlin, Germany, Jun 2728, 2013.
PLANNING UNDER UNCERTAINTY FOR LARGESCALE ACTIVE MULTICAMERA SURVEILLANCE
PROJECT DURATION : Mar 2010  Nov 2014
PROJECT AFFILIATION :
SensorEnhanced Social Media (SeSaMe) Centre (Collaborator: Mohan Kankanhalli)
MEDIA NEWS :
TODAY's Science Section (15 May 2015)  'When CCTV cameras work together as one'
PROBLEM MOTIVATION
The problem of surveillance has grown to be a critical concern in many urban cities worldwide following a recent series of security threats like Mumbai terrorist attacks and London bomb blasts. Central to the problem of surveillance is that of monitoring, tracking, and observing multiple mobile targets of interest distributed over a largescale obstacleridden environment (e.g., airport terminals, railway and subway stations, bus depots, shopping malls, school campuses, military bases). It is often necessary to acquire highresolution videos/images of these targets for supporting realworld surveillance applications like activity/intention tracking and recognition, biometric analysis like target identification and face recognition, surveillance video mining, forensic video analysis/retrieval, among others.
Traditional surveillance systems consist of a large network of fixed/static CCTV (Closed Circuit Television) cameras that are placed to constantly focus at selected important locations in the buildings like entrance/exit, lobby, etc. Unfortunately, the maximum resolution of these cameras is limited to 720 x 480 pixels. So, they cannot capture highresolution images/videos of the targets, especially when the targets are far away from the cameras. As a result, they perform poorly in acquiring the closeup views of the targets and their activities. HDTV/Megapixel cameras have recently been introduced to overcome this resolution issue. Similar to CCTV cameras, these fixed/static HDTV/megapixel cameras are placed to constantly focus at specific locations in the environment. A relatively large network of such cameras has to be installed in order to observe the targets in any region of the environment at high resolution, which is impractical in terms of equipment, installation, and maintenance costs.
The use of active PTZ (Pan/Tilt/Zoom) cameras is becoming an increasingly popular alternative to that of fixed/static cameras for surveillance because the active cameras are endowed with pantiltzoom capabilities that can be exploited to focus on and observe the targets at high image/video resolution. Hence, fewer active cameras need to be deployed to be able to capture highresolution images/videos of the targets in any region of the environment. In order to achieve effective realtime surveillance, an efficient automated mechanism is required to autonomously coordinate and control these cameras' actions.
The objective of this work is thus to address the following central surveillance problem: "How can a network of active cameras be coordinated and controlled to maximize the number of targets observed with a guaranteed image resolution?"
PROPOSED METHODOLOGY
This work presents a novel principled decisiontheoretic planning under uncertainty approach to coordinating and controlling a largescale network of active cameras for performing highquality surveillance of large crowds of moving targets. In particular, our approach addresses the following practical issues affecting the surveillance problem:
(a) Multiple sources of uncertainty. A typical surveillance environment is fraught with multiple sources of uncertainty such as noisy cameras' observations, stochastic targets' motion, and unknown targets' locations, etc. These uncertainties make it difficult for the active cameras to know where to observe in order to keep the targets within their fields of view (fov). Consequently, they may lose track of the observed targets. To resolve this, our approach models a belief over the targets' states (i.e., locations, directions, and velocities) and updates the belief in a Bayesian paradigm based on probabilistic models of the targets' motion and the active cameras' observations;
(b) Cameratarget ratio. In crowded environments, the number of targets to be observed is usually much greater than the number of available cameras. When the number of targets increases, a surveillance system, if poorly designed, tends to incur exponentially increasing computation time, which degrades the realtime performance of the entire surveillance system;
(c) Tradeoff between maximizing the expected number of observed targets and the image resolution of observing them. Increasing the resolution of observing some targets through panning, tilting, or zooming may result in the loss of other targets being tracked. To address this tradeoff, the cameras' actions are coordinated to simultaneously improve the belief over the targets' states and maximize the expected number of targets observed with a guaranteed predefined resolution;
(d) Scalability. By exploiting the inherent structure of the surveillance problem, our approach can scale linearly in the number of targets to be observed;
(e) Realtime requirement. The cameras' actions are computed in real time;
(f) Occlusions. Many realworld surveillance environments contain obstacles that occlude the fov of some or perhaps even all of the cameras, thus preventing the cameras from persistently tracking their observed targets. The regions where the targets cannot be observed by any of the cameras due to obstacles are said to be occluded. When the targets reside in these occluded regions or are not within the fov of any camera, the surveillance system loses track of them, thus degrading the surveillance performance. Such environments are called partially observable in the sense that the exact locations of the targets may not be observed directly by the cameras at all times.
As demonstrated empirically through simulations, our approach can achieve highquality surveillance of a large number of targets in real time and its surveillance performance degrades gracefully with an increasing number of targets. The realworld experiments show the practicality of our decisiontheoretic approach to coordinate and control cameras in surveillance systems.
PUBLICATIONS
 Scalable DecisionTheoretic Coordination and Control for Realtime Active MultiCamera Surveillance.
Prabhu Natarajan, Trong Nghia Hoang, Yongkang Wong, Kian Hsiang Low & Mohan Kankanhalli.
In Proceedings of the
8th ACM/IEEE International Conference on Distributed Smart Cameras (ICDSC'14) (Invited Paper to Special Session on Smart Cameras for Smart Environments), pages 115120, Venezia, Italy, Nov 47, 2014.
Abstract. This paper presents an overview of our novel decisiontheoretic multiagent approach for controlling and coordinating multiple active cameras in surveillance. In this approach, a surveillance task is modeled as a stochastic optimization problem, where the active cameras are controlled and coordinated to achieve the desired surveillance goal in presence of uncertainties. We enumerate the practical issues in active camera surveillance and discuss how these issues are addressed in our decisiontheoretic approach. We focus on two novel surveillance tasks: maximize the number of targets observed in active cameras with guaranteed image resolution and to improve the fairness in observation of multiple targets. We discuss the overview of our novel decisiontheoretic frameworks: Markov decision process and partially observable Markov decision process frameworks for coordinating active cameras in uncertain and partially occluded environments.
 No One is Left "Unwatched": Fairness in Observation of Crowds of Mobile Targets in Active Camera Surveillance.
Prabhu Natarajan, Kian Hsiang Low & Mohan Kankanhalli.
In Proceedings of the
21st European Conference on Artificial Intelligence (ECAI14), including Prestigious Applications of Intelligent Systems (PAIS14), pages 11551160, Prague, Czech Republic, Aug 1822, 2014.
Abstract. Central to the problem of active multicamera surveillance is the fundamental issue of fairness in the observation of crowds of targets such that no target is "starved" of observation by the cameras for a long time. This paper presents a principled decisiontheoretic multicamera coordination and control (MC^{2}) algorithm called fairMC^{2} that can coordinate and control the active cameras to achieve maxmin fairness in the observation of crowds of targets moving stochastically. Our fairMC^{2} algorithm is novel in demonstrating how (a) the uncertainty in the locations, directions, speeds, and observation times of the targets arising from the stochasticity of their motion can be modeled probabilistically, (b) the notion of fairness in observing targets can be formally realized in the domain of multicamera surveillance for the first time by exploiting the maxmin fairness metric to formalize our surveillance objective, that is, to maximize the expected minimum observation time over all targets while guaranteeing a predefined image resolution of observing them, and (c) a structural assumption in the state transition dynamics of a surveillance environment can be exploited to improve its scalability to linear time in the number of targets to be observed during surveillance. Empirical evaluation through extensive simulations in realistic surveillance environments shows that fairMC^{2} outperforms the stateoftheart and baseline MC^{2} algorithms. We have also demonstrated the feasibility of deploying our fairMC^{2} algorithm on real AXIS 214 PTZ cameras.
 DecisionTheoretic Approach to Maximizing Fairness in MultiTarget Observation in MultiCamera Surveillance.
Prabhu Natarajan, Kian Hsiang Low & Mohan Kankanhalli.
In Proceedings of the
13th International Conference on Autonomous Agents and MultiAgent Systems (AAMAS14), pages 15211522, Paris, France, May 59, 2014.
Abstract. Central to the problem of active multicamera surveillance is the fundamental issue of fairness in the observation of multiple targets such that no target is left unobserved by the cameras for a long time. To address this important issue, we propose a novel principled decisiontheoretic approach to control and coordinate multiple active cameras to achieve fairness in the observation of multiple moving targets.
 A DecisionTheoretic Approach for Controlling and Coordinating Multiple Active Cameras in Surveillance.
Prabhu Natarajan.
Ph.D. Thesis, Department of Computer Science, National University of Singapore, Dec 2013.
Abstract.
The use of active cameras in surveillance is becoming increasingly popular as they try to meet the demands of capturing highresolution images/videos of targets in surveillance for face recognition, target identification, forensic video analysis, etc. These active cameras are endowed with pan, tilt, and zoom capabilities, which can be exploited to provide highquality surveillance. In order to achieve effective, realtime surveillance, an efficient collaborative mechanism is needed to control and coordinate these cameras' actions, which is the focus of this thesis. The central problem in surveillance is to monitor a set of targets with guaranteed image resolution. Controlling and coordinating multiple active cameras to achieve this surveillance task is nontrivial and challenging because: (a) presence of inherent uncertainties in the surveillance environment (targets motion, location, and noisy camera observation); (b) there exists a nontrivial tradeoff between number of targets and the resolution of observing these targets; and (c) more importantly, the coordination framework should be scalable with increasing number of targets and cameras.
In this thesis, we formulate a novel decisiontheoretic multiagent planning approach for controlling and coordinating multiple active cameras in surveillance. Our decisiontheoretic approach offers advantages of (a) accounting the uncertainties using probabilistic models; (b) the nontrivial tradeoff is addressed by coordinating the active cameras' actions to maximize the number of targets with guaranteed resolution; and (c) the scalability in number of targets and cameras is achieved by exploiting the structures and properties that are present in our surveillance problem. We focus on two novel problems in active camera surveillance: (a) maximizing observations of multiple targets (MOMT), i.e., maximizing the number of targets observed in active cameras with guaranteed image resolution; and (b) improving fairness in observation of multiple targets (FOMT), i.e., no target is "starved" of observation by active cameras for long duration of time.
We propose two formal decisiontheoretic frameworks (a) Markov Decision Process (MDP) and (b) Partially Observable Markov Decision Process (POMDP) frameworks for coordinating active cameras in surveillance. MDP framework controls active cameras in fully observable surveillance environments where the active cameras are supported by one or more wideview static/fixed cameras to observe the entire surveillance environment at lowresolution. POMDP framework controls active cameras in partially observable surveillance environments where it is impractical to observe the entire surveillance environment using static/fixed cameras due to occlusions caused by physical infrastructures. Hence the POMDP framework do not have a complete view of the surveillance environment.
Specifically, we propose (a) MDP frameworks to solve MOMT problem and FOMT problem in fully observable surveillance environment; and (b) POMDP framework to solve MOMT problem in partially observable surveillance environment. As proven analytically, our MDP and POMDP frameworks incurs time that is linear in number of targets to be observed during surveillance. We have used maxplus algorithm with our MDP framework to improve its scalability in number of cameras for MOMT problem. Empirical evaluation through simulations in realistic surveillance environment reveals that our proposed approach can achieve highquality surveillance in real time. We also demonstrate our pro posed approach with real Axis 214 PTZ cameras to show the practicality of our approach in real world surveillance. Both the simulations and real camera experiments show that our decisiontheoretic approach can control and coordinate active cameras efficiently and hence contributes significantly towards improving the active camera surveillance research.
 DecisionTheoretic Coordination and Control for Active MultiCamera Surveillance in Uncertain, Partially Observable Environments.
Prabhu Natarajan, Trong Nghia Hoang, Kian Hsiang Low & Mohan Kankanhalli.
In Proceedings of the
6th ACM/IEEE International Conference on Distributed Smart Cameras (ICDSC'12), pages 16, Hong Kong, Oct 30  Nov 2, 2012.
Abstract. A central problem of surveillance is to monitor multiple targets moving in a largescale, obstacleridden environment with occlusions. This paper presents a novel principled Partially Observable Markov Decision Processbased approach to coordinating and controlling a network of active cameras for tracking and observing multiple mobile targets at high resolution in such surveillance environments. Our proposed approach is capable of (a) maintaining a belief over the targets' states (i.e., locations, directions, and velocities) to track them, even when they may not be observed directly by the cameras at all times, (b) coordinating the cameras' actions to simultaneously improve the belief over the targets' states and maximize the expected number of targets observed with a guaranteed resolution, and (c) exploiting the inherent structure of our surveillance problem to improve its scalability (i.e., linear time) in the number of targets to be observed. Quantitative comparisons with stateoftheart multicamera coordination and control techniques show that our approach can achieve higher surveillance quality in real time. The practical feasibility of our approach is also demonstrated using real AXIS 214 PTZ cameras.
 PhD Forum: DecisionTheoretic Coordination and Control for Active MultiCamera Surveillance.
Prabhu Natarajan.
In Proceedings of the
6th ACM/IEEE International Conference on Distributed Smart Cameras (ICDSC'12), pages 12, Hong Kong, Oct 30  Nov 2, 2012.
Best PhD Forum Paper Award
Abstract. In this thesis, we present novel decisiontheoretic
multiagent approaches for controlling and coordinating multiple
active cameras in surveillance. Decisiontheoretic approaches
models the interaction between active camera network and the
uncertain surveillance environment effectively. The goal of the
surveillance is to maximize the number of targets observed in
active cameras with guaranteed image resolution. We enumerate
the practical issues in active camera surveillance and discuss how
these issues are addressed in our decisiontheoretic approaches.
The existing camera control approaches have serious limitations
in terms of scalability in number of targets. Where as in
our approaches, the scalability in number of targets has been
improved by exploiting the structure and properties that are
present in our surveillance problem. We proposed two novel
decisiontheoretic frameworks: Markov Decision Process (MDP)
and Partially Observable Markov Decision Process (POMDP)
frameworks for coordinating active cameras in fully observable
and partially observable surveillance settings.
 DecisionTheoretic Approach to Maximizing Observation of Multiple Targets in MultiCamera Surveillance.
Prabhu Natarajan, Trong Nghia Hoang, Kian Hsiang Low & Mohan Kankanhalli.
In Proceedings of the
11th International Conference on Autonomous Agents and MultiAgent Systems (AAMAS12), pages 155162, Valencia, Spain, June 48, 2012.
20.4% acceptance rate
Abstract. This paper presents a novel decisiontheoretic approach to control and coordinate multiple active cameras for observing a number of moving targets in a surveillance system. This approach offers the advantages of being able to (a) account for the stochasticity of targets' motion via probabilistic modeling, and (b) address the tradeoff between maximizing the expected number of observed targets and the resolution of the observed targets through stochastic optimization. One of the key issues faced by existing approaches in multicamera surveillance is that of scalability with increasing number of targets. We show how its scalability can be improved by exploiting the problem structure: as proven analytically, our decisiontheoretic approach incurs time that is linear in the number of targets to be observed during surveillance. As demonstrated empirically through simulations, our proposed approach can achieve highquality surveillance of up to 50 targets in real time and its surveillance performance degrades gracefully with increasing number of targets. We also demonstrate our proposed approach with real AXIS 214 PTZ cameras in maximizing the number of Lego robots observed at high resolution over a surveyed rectangular area. The results are promising and clearly show the feasibility of our decisiontheoretic approach in controlling and coordinating the active cameras in real surveillance system.
 DecisionTheoretic Approach for Controlling and
Coordinating Multiple Active Cameras in Surveillance.
Prabhu Natarajan.
In Proceedings of the
11th International Conference on Autonomous Agents and MultiAgent Systems (AAMAS12), Valencia, Spain, June 48, 2012.
Doctoral consortium abstract
Abstract. This paper presents a novel decisiontheoretic approach to control and coordinate multiple active cameras for observing a number of moving targets in a surveillance system. This approach offers the advantages of being able to (a) account for the stochasticity of targets' motion via probabilistic modeling, and (b) address the tradeoff between maximizing the expected number of observed targets and the resolution of the observed targets through stochastic optimization. One of the key issues faced by existing approaches in multicamera surveillance is that of scalability with increasing number of targets. We show how its scalability can be improved by exploiting the problem structure: as proven analytically, our decisiontheoretic approach incurs time that is linear in the number of targets to be observed during surveillance. As demonstrated empirically through simulations, our proposed approach can achieve highquality surveillance of up to 50 targets in real time and its surveillance performance degrades gracefully with increasing number of targets. We also demonstrate our proposed approach with real AXIS 214 PTZ cameras in maximizing the number of Lego robots observed at high resolution over a surveyed rectangular area. The results are promising and clearly show the feasibility of our decisiontheoretic approach in controlling and coordinating the active cameras in real surveillance system.
MULTIROBOT INFORMATIVE PATH PLANNING FOR ACTIVE SENSING OF SPATIOTEMPORAL ENVIRONMENTAL PHENOMENA
PROJECT DURATION : Jan 2010  May 2013
PROJECT AFFILIATION :
Collaborative MultiRobot Exploration of the Coastal Ocean (Collaborators: John M. Dolan, CMU; Gaurav S. Sukhatme, USC; Kanna Rajan, MBARI)
PROJECT FUNDING : MOE AcRF Tier 1 Grant :
Active Robotic Exploration and Mapping for Environmental Sensing Applications,
SGD $165,377, Apr 2010  Mar 2013
PROBLEM MOTIVATION
Research in environmental sensing and monitoring has recently gained significant attention and practical interest, especially in supporting environmental sustainability efforts worldwide. A key direction of this research aims at sensing, modeling, and predicting the various types of environmental phenomena spatially distributed over our natural and builtup habitats so as to improve our knowledge and understanding of their economic, environmental, and health impacts and implications. This is nontrivial to achieve due to a tradeoff between the quantity of sensing resources (e.g., number of deployed sensors, energy consumption, mission time) and the uncertainty in predictive modeling. In the case of deploying a limited number of mobile robotic sensing assets, such a tradeoff motivates the need to plan the most informative resourceconstrained observation paths to minimize the uncertainty in modeling and predicting a spatially varying environmental phenomenon, which constitutes the active sensing problem to be addressed in this work.
A wide multitude of natural and urban environmental phenomena is characterized by spatially correlated field measurements, which raises the following fundamental issue faced by the active sensing problem:
How can the spatial correlation structure of an environmental phenomenon be exploited to improve the active sensing performance and computational efficiency of robotic path planning?
In this work, we will investigate the above issue for an important broad class of environmental phenomena called anisotropic fields that exhibit a (often much) higher spatial correlation along one direction than along its per pendicular direction. Such fields occur widely in natural and builtup environments and some of them include (a) ocean and freshwater phenomena like plankton density, fish abundance, temperature and salinity; (b) soil and atmospheric phenomena like peat thickness, surface soil moisture, rainfall; (c) mineral deposits like radioactive ore; (d) pollutant and contaminant concentration like air, heavy metals; and (e) ecological abundance like vegetation density.
PROPOSED METHODOLOGY
This work presents two principled approaches to efficient informationtheoretic path planning based on entropy and mutual information criteria for in situ active sensing of environmental phenomena. In contrast to the existing methods described above, our proposed path planning algorithms are novel in addressing a tradeoff between active sensing performance and computational efficiency. An important practical consequence is that our algorithms can exploit the spatial correlation structure of anisotropic fields to improve time efficiency while preserving nearoptimal active sensing performance.
PUBLICATIONS
 MultiRobot Informative Path Planning for Active Sensing of Environmental Phenomena: A Tale of Two Algorithms.
Nannan Cao, Kian Hsiang Low & John M. Dolan.
In Proceedings of the
12th International Conference on Autonomous Agents and MultiAgent Systems (AAMAS13), pages 714, Saint Paul, MN, May 610, 2013.
22.9% acceptance rate
Abstract. A key problem of robotic environmental sensing and monitoring is that of active sensing: How can a team of robots plan the most informative observation paths to minimize the uncertainty in modeling and predicting an environmental phenomenon? This paper presents two principled approaches to efficient informationtheoretic path planning based on entropy and mutual information criteria for in situ active sensing of an important broad class of widelyoccurring environmental phenomena called anisotropic fields. Our proposed algorithms are novel in addressing a tradeoff between active sensing performance and time efficiency. An important practical consequence is that our algorithms can exploit the spatial correlation structure of Gaussian processbased anisotropic fields to improve time efficiency while preserving nearoptimal active sensing performance. We analyze the time complexity of our algorithms and prove analytically that they scale better than stateoftheart algorithms with increasing planning horizon length. We provide theoretical guarantees on the active sensing performance of our algorithms for a class of exploration tasks called transect sampling, which, in particular, can be improved with longer planning time and/or lower spatial correlation along the transect. Empirical evaluation on realworld anisotropic field data shows that our algorithms can perform better or at least as well as the stateoftheart algorithms while often incurring a few orders of magnitude less computational time, even when the field conditions are less favorable.
 InformationTheoretic MultiRobot Path Planning.
Nannan Cao.
M.Sc. Thesis, Department of Computer Science, National University of Singapore, Sep 2012.
Abstract.
Research in environmental sensing and monitoring is especially important in supporting environmental sustainability efforts worldwide, and has recently attracted significant attention and interest. A key direction of this research lies in modeling and predicting the spatiotemporally varying environmental phenomena. One approach is to use a team of robots to sample the area and model the measurement values at unobserved points. For smoothly varying and hotspot fields, there is some work which has been done to model the fields well. However, there is still a class of common environmental fields called anisotropic fields in which the spatial phenomena are highly correlated along one direction and less correlated along the perpendicular direction. We exploit the environmental structure to improve the sampling performance and time efficiency of planning for anisotropic fields.
In this thesis, we cast the planning problem into a stagewise decisiontheoretic problem. we adopt Gaussian Process to model spatial phenomena. Maximum entropy criterion and maximum mutual information criterion are used to measure the informativeness of the observation paths. It is found that for many GPs, correlation of two points exponentially decreases with the distance between the two points. With this property, for maximum entropy criterion, we propose a polynomialtime approximation algorithm, MEPP, to find the maximum entropy paths. We also provide a theoretical performance guarantee for this algorithm. For maximum mutual information criterion, we propose another polynomialtime approximation algorithm, M2IPP. Similar to the MEPP, a performance guarantee is also provided for this algorithm. We demonstrate the performance advantages of our algorithms on two real data sets. To get lower prediction error, three priciples have also been proposed to select the criterion for different environmental fields.
 Active Markov InformationTheoretic Path Planning for Robotic Environmental Sensing.
Kian Hsiang Low, John M. Dolan & Pradeep Khosla.
In Proceedings of the
10th International Conference on Autonomous Agents and MultiAgent Systems (AAMAS11), pages 753760, Taipei, Taiwan, May 26, 2011.
22.1% acceptance rate
Abstract. Recent research in multirobot exploration and mapping has focused on sampling environmental fields, which are typically modeled using the Gaussian process (GP). Existing informationtheoretic exploration strategies for learning GPbased environmental field maps adopt the nonMarkovian problem structure and consequently scale poorly with the length of history of observations. Hence, it becomes computationally impractical to use these strategies for in situ, realtime active sampling. To ease this computational burden, this paper presents a Markovbased approach to efficient informationtheoretic path planning for active sampling of GPbased fields. We analyze the time complexity of solving the Markovbased path planning problem, and demonstrate analytically that it scales better than that of deriving the nonMarkovian strategies with increasing length of planning horizon. For a class of exploration tasks called the transect sampling task, we provide theoretical guarantees on the active sampling performance of our Markovbased policy, from which ideal environmental field conditions and sampling task settings can be established to limit its performance degradation due to violation of the Markov assumption. Empirical evaluation on realworld temperature and plankton density field data shows that our Markovbased policy can generally achieve active sampling performance comparable to that of the widelyused nonMarkovian greedy policies under less favorable realistic field conditions and task settings while enjoying significant computational gain over them.
ENVIRONMENTAL BOUNDARY TRACKING & ESTIMATION WITH MULTIPLE ROBOTS
PROJECT DURATION : Jan 2010  May 2012
PROJECT COLLABORATORS : John M. Dolan, CMU; Steve Chien, JPL, Caltech
PROJECT FUNDING : MOE AcRF Tier 1 Grant :
Active Robotic Exploration and Mapping for Environmental Sensing Applications,
SGD $165,377, Apr 2010  Mar 2013
PROBLEM MOTIVATION
A fundamental problem in environmental sensing and monitoring is to identify and delineate the hotspot regions in a largescale environmental field. It involves partitioning the area spanned by the field into one class of regions called the hotspot regions in which the field measurements exceed a predefined threshold, and the other class of regions where they do not. Such a problem arises in many realworld applications such as precision agriculture, monitoring of ocean and freshwater phenomena (e.g., plankton bloom), forest ecosystems, rare species, pollution (e.g., oil spill), or contamination (e.g., radiation leak). In these applications, it is necessary to assess the spatial extent and shape of the hotspot regions accurately due to severe economic, environmental, and health implications. In practice, this aim is nontrivial to achieve because the constraints on the sampling assets' resources (e.g., energy consumption, mission time, sensing range) limit the number and coverage of in situ observations over the large field that can be used to infer the hotspot regions. Subject to limited observations, the most informative ones should therefore be selected in order to minimize the uncertainty of estimating the hotspot regions (or, equivalently, classifying/labeling the hotspots) in the large field, which motivates our adaptive sampling work in this work.
Mobile robot teams are particularly desirable for performing the above environmental sensing task because they can actively explore to map the hotspot regions at high resolution. On the other hand, static sensors lack mobility and are therefore not capable of doing this well unless a large quantity is deployed. While research in multirobot exploration and mapping have largely focused on the conventional task of building occupancy grids, some recent efforts are put into the more complex, general task of sampling spatially distributed environmental fields. In contrast to occupancy grids that assume discrete, independent cell occupancies, environmental fields are characterized by continuousvalued, spatially correlated measurements, properties of which cannot be exploited by occupancy grid mapping strategies to select the most informative observation paths. To exploit such properties, exploration strategies for learning environmental field maps have recently been developed and can be classified into two regimes: (a) widearea coverage strategies consider sparsely sampled (i.e., largely unexplored) areas to be of high uncertainty and consequently spread observations evenly across the field; (b) hotspot sampling strategies assume areas of high uncertainty and interest to contain extreme, highlyvarying measurements and hence produce clustered observations.
Formal, principled approaches of exploration have also been devised to simultaneously perform hotspot sampling when a hotspot region is found as well as widearea coverage to search for new hotspot regions in sparsely sampled areas. These strategies optimize their observation paths to minimize the uncertainty (e.g., in terms of meansquared error or entropy) of mapping the entire continuousvalued field. They are, however, suboptimal for classifying/labeling the hotspots in the field, which we will discuss and demonstrate theoretically and empirically in this work.
PROPOSED METHODOLOGY
This work presents a novel decentralized active robotic exploration (DARE) strategy for probabilistic classification/labeling of hotspots in a largescale environmental field. The environmental field is assumed to be realized from a rich class of probabilistic spatial models called Gaussian process (GP) that can formally characterize its spatial correlation structure. More importantly, it can provide formal measures of classification/labeling uncertainty (i.e., in the form of cost functions) such as the misclassification and entropy criteria for directing the robots to explore highly uncertain areas of the field. The chief impediment to using these formal criteria is that they result in costminimizing exploration strategies, which cannot be solved in closed form. To resolve this, they are reformulated as rewardmaximizing dual strategies, from which we can then derive the approximate DARE strategy to be solved in closed form efficiently. The specific contributions of our work include:
 Analyzing the time complexity of solving the DARE strategy: We prove that its incurred time is independent of the map resolution and the number of robots, thus making it practical for in situ, realtime active sampling. In contrast, existing stateoftheart exploration strategies for learning environmental field maps scale poorly with increasing map resolution and/or number of robots;
 Analyzing the exploration behavior of the DARE strategy through its formulation: It exhibits an interesting formal tradeoff between that of boundary tracking until the hotspot region boundary can be accurately predicted and widearea coverage to find new boundaries in sparsely sampled areas to be tracked. In contrast, ad hoc, reactive boundary tracking strategies typically require a hotspot region boundary to be located manually or via random exploration and are not driven by the need to maximize the fidelity of estimating multiple hotspot regions given limited observations;
 Providing theoretical guarantee on the active exploration performance of the DARE strategy: We prove that, under a reasonable conditional independence assumption, it produces the same optimal observation paths as that of the centralized costminimizing strategies, the latter of which otherwise cannot be solved in closed form. This result has a simple but important implication: The uncertainty of labeling the hotspots in a GPbased field is greatest at or close to the hotspot region boundaries;
 Empirically evaluating the active exploration performance and time efficiency of the DARE strategy on realworld plankton density and temperature field data: Subject to limited observations, the DARE strategy can achieve better classification of the hotspots than stateoftheart active exploration strategies while being significantly more timeefficient than those performing widearea coverage and hotspot sampling.
PUBLICATIONS
 Decentralized Active Robotic Exploration and Mapping for Probabilistic Field Classification in Environmental Sensing.
Kian Hsiang Low, Jie Chen, John M. Dolan, Steve Chien & David R. Thompson.
In Proceedings of the
11th International Conference on Autonomous Agents and MultiAgent Systems (AAMAS12), pages 105112, Valencia, Spain, June 48, 2012.
20.4% acceptance rate
Also appeared in
IROS'11 Workshop on Robotics for Environmental Monitoring (WREM11), San Francisco, CA, Sep 30, 2011.
Abstract. A central problem in environmental sensing and monitoring is to classify/label the hotspots in a largescale environmental field. This paper presents a novel decentralized active robotic exploration (DARE) strategy for probabilistic classification/labeling of hotspots in a Gaussian process (GP)based field. In contrast to existing stateoftheart exploration strategies for learning environmental field maps, the time needed to solve the DARE strategy is independent of the map resolution and the number of robots, thus making it practical for in situ, realtime active sampling. Its exploration behavior exhibits an interesting formal tradeoff between that of boundary tracking until the hotspot region boundary can be accurately predicted and widearea coverage to find new boundaries in sparsely sampled areas to be tracked. We provide a theoretical guarantee on the active exploration performance of the DARE strategy: under reasonable conditional independence assumption, we prove that it can optimally achieve two formal costminimizing exploration objectives based on the misclassification and entropy criteria. Importantly, this result implies that the uncertainty of labeling the hotspots in a GPbased field is greatest at or close to the hotspot region boundaries. Empirical evaluation on realworld plankton density and temperature field data shows that, subject to limited observations, DARE strategy can achieve more superior classification of hotspots and time efficiency than stateoftheart active exploration strategies.
MULTIROBOT ADAPTIVE SAMPLING FOR ENVIRONMENTAL SENSING & MONITORING
PROJECT DURATION : Jul 2005  Mar 2010
PROJECT LAB : TSAR
PROBLEM MOTIVATION
Recent research in multirobot exploration and mapping has focused on sampling environmental fields, some of which typically feature a few small hotspots in a large region.
Such a hotspot field often arises in two realworld applications:
(1) planetary exploration such as geologic reconnaissance and prospecting for mineral deposits or natural gases, and
(2) environment and ecological sensing such as precision agriculture, and monitoring of ocean phenomena (e.g., plankton bloom, anoxic zones), forest ecosystems, rare species, pollution (e.g., oil spill), or contamination (e.g., radiation leak).
In particular, the hotspot field is characterized by continuousvalued, spatially correlated measurements with the hotspots exhibiting extreme measurements and much higher spatial variability than the rest of the field.
With limited (e.g., pointbased) robot sensing range, a complete coverage becomes impractical in terms of resource costs (e.g., resource consumption).
So, to accurately map the field, the hotspots have to be sampled at a higher resolution.
The hotspot field discourages static sensor placement because a large number of sensors has to be positioned to detect and refine the sampling of hotspots. If these static sensors are not placed in any hotspot initially, they cannot reposition by themselves to locate one. In contrast, a robot team is capable of performing highresolution hotspot sampling due to its mobility. Hence, it is desirable to build a mobile robot team that can actively explore to map a hotspot field.
PROPOSED METHODOLOGY
To learn a hotspot field map, the exploration strategy of the robot team has to plan the most informative resourceconstrained observation paths that minimize the uncertainty of mapping the hotspot field.
By representing the hotspot field using rich classes of Bayesian nonparametric models such as the Gaussian process or logGaussian process,
formal measures of mapping uncertainty (e.g., based on meansquared error [ AAMAS08] or entropy [ ICAPS09] criterion) can be defined and subsequently exploited by our proposed adaptive sampling algorithms for directing the robot team to explore highly uncertain areas of the field.
In contrast to nonadaptive sampling strategies that only perform well with smoothlyvarying fields,
our nonmyopic adaptive sampling algorithms can exploit clustering phenomena (i.e., hotspots) to plan observation paths that produce lower mapping uncertainty.
PUBLICATIONS
 Adaptive Sampling of Time Series with Application to Remote Exploration.
David R. Thompson, Nathalie Cabrol, Michael Furlong, Craig Hardgrove, Kian Hsiang Low, Jeffrey Moersch & David Wettergreen.
In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA'13), pages 34633468, Karlsruhe, Germany, May 610, 2013.
Abstract. We address the problem of adaptive informationoptimal data collection in time series. Here a remote sensor or explorer agent throttles its sampling rate in order to track anomalous events while obeying constraints on time and power. This problem is challenging because the agent has limited visibility  all collected datapoints lie in the past, but its resource allocation decisions require predicting far into the future. Our solution is to continually fit a Gaussian process model to the latest data and optimize the sampling plan on line to maximize information gain. We compare the performance characteristics of stationary and nonstationary Gaussian process models. We also describe an application based on geologic analysis during planetary rover exploration. Here adaptive sampling can improve coverage of localized anomalies and potentially benefit mission science yield of long autonomous traverses.
 Telesupervised Remote Surface
Water Quality Sensing.
Gregg Podnar, John M. Dolan, Kian Hsiang Low & Alberto Elfes.
In Proceedings of the IEEE Aerospace Conference, Big Sky, MT, Mar 613, 2010.
Abstract. We present a fleet of autonomous Robot Sensor Boats (RSBs) developed for lake and river fresh water quality assessment and controlled by our Multilevel Autonomy Robot Telesupervision Architecture (MARTA). The RSBs are low cost, highly maneuverable, shallow draft sensor boats, developed as part of the Sensor Web program supported under the Advanced Information Systems Technology program of NASA's Earth Systems Technology Office. They can scan large areas of lakes, and navigate up tributaries to measure water quality near outfalls that larger research vessels cannot reach. The MARTA telesupervision architecture has been applied to a number of domains from multiplatform autonomous wide area planetary mineral prospecting, to multiplatform ocean monitoring. The RSBs are a complementary expansion of a fleet of NOAA/NASAdeveloped extendeddeployment surface autonomous vehicles that enable insitu study of meteorological factors of the ocean/atmosphere interface, and which have been adapted to investigate harmful algal blooms under this program. The flexibility of the MARTA telesupervision architecture was proven as it supported simultaneous operation of these heterogenous autonomous sensor platforms while geographically widely separated. Results and analysis are presented of multiple tests carried out over three months using a multisensor water sonde to assess water quality in a small recreational lake. Inference Grids were used to produce maps representing temperature, pH, and dissolved oxygen. The tests were performed under various water conditions (clear vs. hair algaeladen) and both before and after heavy rains. Data from each RSB was relayed to a data server in our lab in Pittsburgh, Pennsylvania, and made available over the World Wide Web where it was acquired by team members at the Jet Propulsion Laboratory of NASA in Pasadena, California who monitored the boats and their sensor readings in real time, as well as using these data to model the water quality by producing Inference Gridbased maps.
 MultiRobot Adaptive Exploration and Mapping for Environmental Sensing Applications.
Kian Hsiang Low.
Ph.D. Thesis, Technical Report CMUECE2009024, Department of Electrical and Computer Engineering, Carnegie Mellon University, Pittsburgh, PA, Aug 2009.
Abstract.
Recent research in robot exploration and mapping has focused on sampling hotspot fields, which often arise in environmental and ecological sensing applications. Such a hotspot field is characterized by continuous, positively skewed, spatially correlated measurements with the hotspots exhibiting extreme measurements and much higher spatial variability than the rest of the field.
To map a hotspot field of the above characterization, we assume that it is realized from nonparametric probabilistic models such as the Gaussian and logGaussian processes (respectively, GP and lGP), which can provide formal measures of map uncertainty. To learn a hotspot field map, the exploration strategy of a robot team then has to plan resourceconstrained observation paths that minimize the uncertainty of a spatial model of the hotspot field. This exploration problem is formalized in a sequential decisiontheoretic planning under uncertainty framework called the multirobot adaptive sampling problem (MASP). So, MASP can be viewed as a sequential, nonmyopic version of active learning. In contrast to finitestate Markov decision problems, MASP adopts a more complex but realistic continuousstate, nonMarkovian problem structure so that its induced exploration policy can be informed by the complete history of continuous, spatially correlated observations for selecting paths. It is unique in unifying formulations of nonmyopic exploration problems along the entire adaptivity spectrum, thus subsuming existing nonadaptive formulations and allowing the performance advantage of a more adaptive policy to be theoretically realized. Through MASP, it is demonstrated that a more adaptive strategy can exploit clustering phenomena in a hotspot field to produce lower expected map uncertainty. By measuring map uncertainty using the meansquared error criterion, a MASPbased exploration strategy consequently plans adaptive observation paths that minimize the expected posterior map error or equivalently, maximize the expected map error reduction.
The time complexity of solving MASP (approximately) depends on the map resolution, which limits its practical use in largescale, highresolution exploration and mapping. This computational difficulty is alleviated through an informationtheoretic approach to MASP (iMASP), which measures map uncertainty based on the entropy criterion instead. As a result, an iMASPbased exploration strategy plans adaptive observation paths that minimize the expected posterior map entropy or equivalently, maximize the expected entropy of observation paths. Unlike MASP, reformulating the costminimizing iMASP as a rewardmaximizing dual problem causes its time complexity of being solved approximately to be independent of the map resolution and less sensitive to larger robot team size as demonstrated both analytically and empirically. Furthermore, this rewardmaximizing dual transforms the widelyused nonadaptive maximum entropy sampling problem into a novel adaptive variant, thus improving the performance of the induced exploration policy.
One advantage stemming from the rewardmaximizing dual formulations of MASP and iMASP is that they allow observation selection properties of the induced exploration policies to be realized for sampling the hotspot field. These properties include adaptivity, hotspot sampling, and widearea coverage. We show that existing GPbased exploration strategies may not explore and map the hotspot field well with the selected observations because they are nonadaptive and perform only widearea coverage. In contrast, the lGPbased exploration policies can learn a highquality hotspot field map because they are adaptive and perform both widearea coverage and hotspot sampling.
The other advantage is that even though MASP and iMASP are nontrivial to solve due to their continuous state components, the convexity of their rewardmaximizing duals can be exploited to derive, in a computationally tractable manner, discretestate monotonebounding approximations and subsequently, approximately optimal exploration policies with theoretical performance guarantees. Anytime algorithms based on approximate MASP and iMASP are then proposed to alleviate the computational difficulty that arises from their nonMarkovian structure.
It is of practical interest to be able to quantitatively characterize the "hotspotness" of an environmental field. We propose a novel "hotspotness" index, which is defined in terms of the spatial correlation properties of the hotspot field. As a result, this index can be related to the intensity, size, and diffuseness of the hotspots in the field.
We also investigate how the spatial correlation properties of the hotspot field affect the performance advantage of adaptivity. In particular, we derive sufficient and necessary conditions of the spatial correlation properties for adaptive exploration to yield no performance advantage.
Lastly, we develop computationally efficient approximately optimal exploration strategies for sampling the GP by assuming the Markov property in iMASP planning. We provide theoretical guarantees on the performance of the Markovbased policies, which improve with decreasing spatial correlation. We evaluate empirically the effects of varying spatial correlations on the mapping performance of the Markovbased policies as well as whether these Markovbased path planners are timeefficient for the transect sampling task.
Through the abovementioned work, this thesis establishes the following two claims: (1) adaptive, nonmyopic exploration strategies can exploit clustering phenomena to plan observation paths that produce lower map uncertainty than nonadaptive, greedy methods; and (2) Markovbased exploration strategies can exploit small spatial correlation to plan observation paths which achieve map uncertainty comparable to that of nonMarkovian policies using significantly less planning time.
 InformationTheoretic Approach to Efficient Adaptive Path Planning for Mobile Robotic Environmental Sensing.
Kian Hsiang Low, John M. Dolan & Pradeep Khosla.
In Proceedings of the 19th International Conference on Automated Planning and Scheduling (ICAPS09), pages 233240, Thessaloniki, Greece, Sep 1923, 2009.
33.9% acceptance rate
Also appeared in IPSN09 Workshop on Sensor Networks for Earth and Space Science Applications (ESSA09), San Francisco, CA, Apr 16, 2009.
Also orally presented in RSS09 Workshop on Aquatic Robots and Ocean Sampling, Seattle, WA, Jun 29, 2009.
Abstract. Recent research in robot exploration and mapping has focused on sampling environmental hotspot fields. This exploration task is formalized by Low, Dolan, and Khosla (2008) in a sequential decisiontheoretic planning under uncertainty framework called MASP. The time complexity of solving MASP approximately depends on the map resolution, which limits its use in largescale, highresolution exploration and mapping. To alleviate this computational difficulty, this paper presents an informationtheoretic approach to MASP (iMASP) for efficient adaptive path planning; by reformulating the costminimizing iMASP as a rewardmaximizing problem, its time complexity becomes independent of map resolution and is less sensitive to increasing robot team size as demonstrated both theoretically and empirically. Using the rewardmaximizing dual, we derive a novel adaptive variant of maximum entropy sampling, thus improving the induced exploration policy performance. It also allows us to establish theoretical bounds quantifying the performance advantage of optimal adaptive over nonadaptive policies and the performance quality of approximately optimal vs. optimal adaptive policies. We show analytically and empirically the superior performance of iMASPbased policies for sampling the logGaussian process to that of policies for the widelyused Gaussian process in mapping the hotspot field. Lastly, we provide sufficient conditions that, when met, guarantee adaptivity has no benefit under an assumed environment model.
 Cooperative Aquatic Sensing using the Telesupervised Adaptive Ocean Sensor Fleet.
John M. Dolan, Gregg W. Podnar, Stephen Stancliff, Kian Hsiang Low, Alberto Elfes, John Higinbotham, Jeffrey C. Hosler, Tiffany A. Moisan & John Moisan.
In Proceedings of the SPIE Conference on Remote Sensing of the Ocean, Sea Ice, and Large Water Regions, volume 7473, Berlin, Germany, Aug 31  Sep 3, 2009.
Abstract. Earth science research must bridge the gap between the atmosphere and the ocean to foster understanding of Earth's climate and ecology. Typical ocean sensing is done with satellites or in situ buoys and research ships which are slow to reposition. Cloud cover inhibits study of localized transient phenomena such as Harmful Algal Blooms (HAB). A fleet of extendeddeployment surface autonomous vehicles will enable in situ study of characteristics of HAB, coastal pollutants, and related phenomena. We have developed a multiplatform telesupervision architecture that supports adaptive reconfiguration based on environmental sensor inputs. Our system allows the autonomous repositioning of smart sensors for HAB study by networking a fleet of NOAA OASIS (Ocean Atmosphere Sensor Integration System) surface autonomous vehicles. In situ measurements intelligently modify the search for areas of high concentration. Inference Grid and complementary informationtheoretic techniques support sensor fusion and analysis. Telesupervision supports sliding autonomy from highlevel mission tasking, through vehicle and data monitoring, to teleoperation when direct human interaction is appropriate. This paper reports on experimental results from multiplatform tests conducted in the Chesapeake Bay and in Pittsburgh, Pennsylvania waters using OASIS platforms, autonomous kayaks, and multiple simulated platforms to conduct cooperative sensing of chlorophylla and water quality.
 Robot Boats as a Mobile Aquatic Sensor Network.
Kian Hsiang Low, Gregg Podnar, Stephen Stancliff, John M. Dolan & Alberto Elfes.
In Proceedings of the IPSN09 Workshop on Sensor Networks for Earth and Space Science Applications (ESSA09), San Francisco, CA, Apr 16, 2009.
Abstract. This paper describes the Multilevel Autonomy Robot Telesupervision Architecture (MARTA), an architecture for supervisory control of a heterogeneous fleet of networked unmanned autonomous aquatic surface vessels carrying a payload of environmental science sensors. This architecture allows a landbased human scientist to effectively supervise data gathering by multiple robotic assets that implement a web of widely dispersed mobile sensors for in situ study of physical, chemical or biological processes in water or in the water/atmosphere interface.
 Adaptive MultiRobot WideArea Exploration And Mapping.
Kian Hsiang Low, John M. Dolan & Pradeep Khosla.
In Proceedings of the
7th International Conference on Autonomous Agents and MultiAgent Systems (AAMAS08), pages 2330, Estoril, Portugal, May 1216, 2008.
22.2% acceptance rate
Also presented as a poster in RSS09 Workshop on Aquatic Robots and Ocean Sampling, Seattle, WA, Jun 29, 2009.
Abstract. The exploration problem is a central issue in mobile robotics. A complete terrain coverage is not practical if the environment is large with only a few small hotspots. This paper presents an adaptive multirobot exploration strategy that is novel in performing both widearea coverage and hotspot sampling using nonmyopic path planning. As a result, the environmental phenomena can be accurately mapped. It is based on a dynamic programming formulation, which we call the Multirobot Adaptive Sampling Problem (MASP). A key feature of MASP is in covering the entire adaptivity spectrum, thus allowing strategies of varying adaptivity to be formed and theoretically analyzed in their performance; a more adaptive strategy improves mapping accuracy. We apply MASP to sampling the Gaussian and logGaussian processes, and analyze if the resulting strategies are adaptive and maximize widearea coverage and hotspot sampling. Solving MASP is nontrivial as it comprises continuous state components. So, it is reformulated for convex analysis, which allows discretestate monotonebounding approximation to be developed. We provide a theoretical guarantee on the policy quality of the approximate MASP (aMASP) for using in MASP. Although aMASP can be solved exactly, its state size grows exponentially with the number of stages. To alleviate this computational difficulty, anytime algorithms are proposed based on aMASP, one of which can guarantee its policy quality for MASP in real time.
 Adaptive Sampling for MultiRobot WideArea Exploration.
Kian Hsiang Low, Geoffrey J. Gordon, John M. Dolan & Pradeep Khosla.
In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA'07), pages 755760, Rome, Italy, Apr 1014, 2007.
Abstract. The exploration problem is a central issue in mobile robotics. A complete coverage is not practical if the environment is large with a few small hotspots, and the sampling cost is high. So, it is desirable to build robot teams that can coordinate to maximize sampling at these hotspots while minimizing resource costs, and consequently learn more accurately about properties of such environmental phenomena. An important issue in designing such teams is the exploration strategy. The contribution of this paper is in the evaluation of an adaptive exploration strategy called adaptive cluster sampling (ACS), which is demonstrated to reduce the resource costs (i.e., mission time and energy consumption) of a robot team, and yield more information about the environment by directing robot exploration towards hotspots. Due to the adaptive nature of the strategy, it is not obvious how the sampled data can be used to provide unbiased, lowvariance estimates of the properties. This paper therefore discusses how estimators that are RaoBlackwellized can be used to achieve low error. This paper also presents the first analysis of the characteristics of the environmental phenomena that favor the ACS strategy and estimators. Quantitative experimental results in a mineral prospecting task simulation show that our approach is more efficient in exploration by yielding more minerals and information with fewer resources and providing more precise mineral density estimates than previous methods.
 Adaptive Sampling for MultiRobot Wide Area Prospecting.
Kian Hsiang Low, Geoffrey J. Gordon, John M. Dolan, and Pradeep Khosla.
In Technical Report CMURITR0551, Robotics Institute, Carnegie Mellon University, Pittsburgh, PA, Oct 2005.
Abstract. Prospecting for in situ mineral resources is essential for establishing settlements on the Moon and Mars. To reduce human effort and risk, it is desirable to build robotic systems to perform this prospecting. An important issue in designing such systems is the sampling strategy: how do the robots choose where to prospect next? This paper argues that a strategy called Adaptive Cluster Sampling (ACS) has a number of desirable properties: compared to conventional strategies, (1) it reduces the total mission time and energy consumption of a team of robots, and (2) returns a higher mineral yield and more information about the prospected region by directing exploration towards areas of high mineral density, thus providing detailed maps of the boundaries of such areas. Due to the adaptive nature of the sampling scheme, it is not immediately obvious how the resulting sampled data can be used to provide an unbiased, lowvariance estimate of the regional mineral density. This paper therefore investigates new mineral density estimators, which have lower error than previouslydeveloped estimators; they are derived from the older estimators via a process called RaoBlackwellization. Since the efficiency of estimators depends on the type of mineralogical population sampled, the population characteristics that favor ACS estimators are also analyzed. The ACS scheme and our new estimators are evaluated empirically in a detailed simulation of the prospecting task, and the quantitative results show that our approach can yield more minerals with less resources and provide more accurate mineral density estimates than previous methods.
DISTRIBUTED LAYERED ARCHITECTURE FOR SELFORGANIZING MOBILE SENSOR NETWORKS
PROJECT DURATION : Nov 2002  Jun 2005
PROBLEM MOTIVATION
One of the fundamental issues that arises in a sensor network is coverage.
Traditionally, network coverage is maximized by determining the optimal placement of static sensors in a centralized manner,
which can be related to the class of art gallery problems.
However, recent investigations in sensor network mobility reveal that
mobile sensors can selforganize to provide better coverage than static placement.
Existing applications have only utilized uninformed mobility (i.e., random motion or patrol).
In contrast, our work here focuses on informed, intelligent mobility to further improve coverage.
Our network coverage problem is motivated by the following constraints
that discourage static sensor placement or uninformed mobility:
(a) no prior information about the exact target locations, population densities or motion pattern,
(b) limited sensory range, and
(c) very large area to be observed.
All these conditions may cause the sensors to be unable to cover the entire region of interest.
Hence, fixed sensor locations or uninformed mobility will not be adequate in general.
Rather, the sensors have to move dynamically in response to the motion and distribution of targets and other sensors
to maximize coverage.
Inspired by robotics, the above problem may be regarded as that of lowlevel motion control to coordinate the sensors'
target tracking movements in the continuous workspace.
Alternatively, it can be cast as a highlevel task allocation problem by segmenting the workspace into discrete regions
such that each region is assigned a group or coalition of sensors to track the targets within.
PROPOSED METHODOLOGY
This work presents a reactive layered multirobot architecture
for distributed mobile sensor network coverage in complex, dynamic environments.
At the lower layer, each robot uses a reactive motion control strategy
known as Cooperative Extended Kohonen Maps
to coordinate their target tracking within a region without the need of communication.
This strategy is also responsible for obstacle avoidance, robot separation to minimize task interference,
and navigation between regions via beacons or checkpoints plotted by a motion planner.
At the higher layer, the robots use a dynamic antbased task allocation scheme to cooperatively selforganize
their coalitions in a decentralized manner according to the target distributions across the regions.
This scheme addresses the following issues, which distinguish it from the other task allocation mechanisms:
Task Allocation for MultiRobot Tasks: Existing algorithms (e.g., auctionand behaviorbased) assume a multirobot
task can be partitioned into singlerobot tasks. But this may not be always possible or the multirobot task
can be more efficiently performed by coalitions of robots.
Coalition Formation for Minimalist Robots: Existing coalition formation schemes require complex planning, explicit
negotiation, and precise estimation of coalitional cost. Hence, they do not perform well in dynamic, realtime scenarios.
Cooperation of ResourceLimited Robots: Robots with limited communication and sensing capabilities (i.e., partial
observability) can only obtain local, uncertain information of the dynamic environment. With limited computational power,
their cooperative strategies cannot involve complex planning or negotiations.
PUBLICATIONS
 Autonomic Mobile Sensor Network with SelfCoordinated Task Allocation and Execution.
Kian Hsiang Low, Wee Kheng Leow & Marcelo H. Ang, Jr.
IEEE Transactions on Systems, Man, and Cybernetics  Part C: Applications and Reviews
(Special Issue on Engineering Autonomic Systems), volume 36, issue 3, pages 315327, May 2006.
Extended version of our ICRA'04 and
AAAI04 papers
Andrew P. Sage Best Transactions Paper Award for the best paper published in IEEE Trans. SMC  Part A, B, and C in 2006
Abstract. This paper describes a distributed layered architecture for resourceconstrained multirobot cooperation, which is utilized in autonomic mobile sensor network coverage. In the upper layer, a dynamic task allocation scheme selforganizes the robot coalitions to track efficiently across regions. It uses concepts of ant behavior to selfregulate the regional distributions of robots in proportion to that of the moving targets to be tracked in a nonstationary environment. As a result, the adverse effects of task interference between robots are minimized and network coverage is improved. In the lower task execution layer, the robots use selforganizing neural networks to coordinate their target tracking within a region. Both layers employ selforganization techniques, which exhibit autonomic properties such as selfconfiguring, selfoptimizing, selfhealing, and selfprotecting. Quantitative comparisons with other tracking strategies such as static sensor placements, potential fields, and auctionbased negotiation show that our layered approach can provide better coverage, greater robustness to sensor failures, and greater flexibility to respond to environmental changes.
 Task Allocation via SelfOrganizing Swarm Coalitions in Distributed Mobile Sensor Network.
Kian Hsiang Low, Wee Kheng Leow & Marcelo H. Ang, Jr.
In Proceedings of the 19th National Conference on Artificial Intelligence (AAAI04), pages 2833, San Jose, CA, Jul 2529, 2004.
26.7% acceptance rate
Abstract. This paper presents a task allocation scheme via selforganizing swarm coalitions for distributed mobile sensor network coverage. Our approach uses the concepts of ant behavior to selfregulate the regional distributions of sensors in proportion to that of the moving targets to be tracked in a nonstationary environment. As a result, the adverse effects of task interference between robots are minimized and sensor network coverage is improved. Quantitative comparisons with other tracking strategies such as static sensor placement, potential fields, and auctionbased negotiation show that our approach can provide better coverage and greater flexibility to respond to environmental changes.
 Reactive, Distributed Layered Architecture for ResourceBounded MultiRobot Cooperation: Application to Mobile Sensor Network Coverage.
Kian Hsiang Low, Wee Kheng Leow & Marcelo H. Ang, Jr.
In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA'04), pages 37473752, New Orleans, LA, Apr 26  May 1, 2004.
Abstract. This paper describes a reactive, distributed layered architecture for cooperation of multiple resourcebounded robots, which is utilized in mobile sensor network coverage. In the upper layer, a dynamic task allocation scheme selforganizes the robot coalitions to track efficiently in separate regions. It uses the concepts of ant behavior to selfregulate the regional distributions of robots in proportion to that of the targets to be tracked in the changing environment. As a result, the adverse effects of task interference between robots are minimized and sensor network coverage is improved. In the lower layer, the robots use selforganizing neural networks to coordinate their target tracking within a region. Quantitative comparisons with other tracking strategies such as static sensor placements, potential fields, and auctionbased negotiation show that our approach can provide better coverage and greater flexibility in responding to environmental changes.
PRESENTATIONS
 Task Allocation via SelfOrganizing Swarm Coalitions in Distributed Mobile Sensor Network.
Kian Hsiang Low.
Presented in 8th National IT Awareness Project Competition (NITA04), National University of Singapore, Mar 13, 2004
(Overall Best Project, Postgraduate Category).
VIDEO DEMOS
Coverage of 30 targets (green) with 15 ant robots (white)
 Selforganization of swarm coalitions to unknown, timevarying target distribution after
 Robot switching to region of higher task demand (i.e., targets to robots ratio)
ACTION SELECTION MECHANISM FOR MULTIROBOT TASKS
PROJECT DURATION : Sep 2002  Nov 2002
PROBLEM MOTIVATION
A central issue in the design of behaviorbased control architectures for autonomous mobile robots
is the formulation of effective mechanisms to coordinate the behaviors.
These mechanisms determine the policy of conflict resolution between behaviors,
which involves behavioral cooperation and competition to select the most appropriate action.
The actions are selected so as to optimize the achievement of the goals or behavioral objectives.
Developing such an action selection methodology is nontrivial
due to realistic constraints such as environmental complexity and unpredictability,
and resource limitations, which include computational and cognitive capabilities of the robot,
incomplete knowledge of the environment, and time constraints.
As a result, action selection can never be absolutely optimal.
Given these constraints, the action selection scheme should be able to choose actions
that are good enough to satisfy multiple concurrent, possibly conflicting, behavioral objectives.
PROPOSED METHODOLOGY
Our motivation of the action selection mechanism is to develop a motion control strategy for autonomous nonholonomic mobile robots
that can perform distributed multirobot surveillance in unknown, dynamic, complex, and unpredictable environments.
By implementing the action selection framework using an assemblage of selforganizing neural networks,
it induces the following key features that significantly enhance the agent's action selection capability:
selforganization of continuous state and action spaces to provide smooth, efficient and fine motion control,
and action selection via the cooperation and competition of Extended Kohonen Maps to achieve more complex motion tasks:
(1) negotiation of unforeseen concave and narrowly spaced obstacles, and
(2) cooperative tracking of multiple mobile targets by a team of robots.
Qualitative and quantitative comparisons for single and multirobot tasks show that
our framework can provide better action selection than do potential fields method.
PUBLICATIONS
 An Ensemble of Cooperative Extended Kohonen Maps for Complex Robot Motion Tasks.
Kian Hsiang Low, Wee Kheng Leow & Marcelo H. Ang, Jr.
Neural Computation, volume 17, issue 6, pages 14111445, Jun 2005.
Extended version of our
IJCAI03 paper
Abstract. Selforganizing feature maps such as extended Kohonen maps (EKMs) have been very successful at learning sensorimotor control for mobile robot tasks. This letter presents a new ensemble approach, cooperative EKMs with indirect mapping, to achieve complex robot motion. An indirectmapping EKM selforganizes to map from the sensory input space to the motor control space indirectly via a control parameter space. Quantitative evaluation reveals that indirect mapping can provide finer, smoother, and more efficient motion control than does direct mapping by operating in a continuous, rather than discrete, motor control space. It is also shown to outperform basis function neural networks. Furthermore, training its control parameters with recursive least squares enables faster convergence and better performance compared to gradient descent. The cooperation and competition of multiple selforganized EKMs allow a nonholonomic mobile robot to negotiate unforeseen, concave, closely spaced, and dynamic obstacles. Qualitative and quantitative comparisons with neural network ensembles employing weighted sum reveal that our method can achieve more sophisticated motion tasks even though the weightedsum ensemble approach also operates in continuous motor control space.
 ContinuousSpaced Action Selection for Single and MultiRobot Tasks Using Cooperative Extended Kohonen Maps.
Kian Hsiang Low, Wee Kheng Leow & Marcelo H. Ang, Jr.
In Proceedings of the IEEE International Conference on Networking, Sensing and Control (ICNSC'04)
(Invited Paper to Special Session on Visual Surveillance), pages 198203, Taipei, Taiwan, Mar 2123, 2004.
Abstract. Action selection is a central issue in the design of behaviorbased control architectures for autonomous mobile robots. This paper presents an action selection framework based on an assemblage of selforganizing neural networks called Cooperative Extended Kohonen Maps. This framework encapsulates two features that significantly enhance a robot's action selection capability: selforganization in the continuous state and action spaces to provide smooth, efficient and fine motion control; action selection via the cooperation and competition of Extended Kohonen Maps so that more complex motion tasks can be achieved. Qualitative and quantitative comparisons for both single and multirobot motion tasks show that our framework can provide better action selection than do action superposition methods.
 Action Selection for Single and MultiRobot Tasks Using Cooperative Extended Kohonen Maps.
Kian Hsiang Low, Wee Kheng Leow & Marcelo H. Ang, Jr.
In Proceedings of the 18th International Joint Conference on Artificial Intelligence (IJCAI03), pages 15051506, Acapulco, Mexico, Aug 915, 2003.
27.6% acceptance rate
Abstract. This paper presents an action selection framework based on an assemblage of selforganizing neural networks called Cooperative Extended Kohonen Maps. This framework encapsulates two features that significantly enhance a robot's action selection capability: selforganization in the continuous state and action spaces to provide smooth, efficient and fine motion control; action selection via the cooperation and competition of Extended Kohonen Maps to achieve more complex motion tasks. Qualitative and quantitative comparisons for single and multirobot tasks show our framework can provide better action selection than do potential fields method.
 Action Selection in Continuous State and Action Spaces by Cooperation and Competition of Extended Kohonen Maps.
Kian Hsiang Low, Wee Kheng Leow & Marcelo H. Ang, Jr.
In Proceedings of the
2nd International Joint Conference on Autonomous Agents and MultiAgent Systems (AAMAS03), pages 10561057, Melbourne, Australia, Jul 1418, 2003.
Abstract. This paper presents an action selection framework based on an assemblage of selforganizing neural networks called Cooperative Extended Kohonen Maps. This framework encapsulates two features that significantly enhance a robot's action selection capability: selforganization in the continuous state and action spaces to provide smooth, efficient and fine motion control; action selection via the cooperation and competition of Extended Kohonen Maps to achieve more complex motion tasks. Qualitative tests demonstrate the capability of our action selection method for both single and multirobot motion tasks.
VIDEO DEMO
 Cooperative tracking of moving targets by robots
using cooperative Extended Kohonen Maps
INTEGRATED ROBOT PLANNING AND CONTROL
PROJECT DURATION : Jul 2001  Sep 2002
PROBLEM MOTIVATION
Robot motion research has proceeded along two separate directions:
highlevel deliberative planning and lowlevel reactive control.
Deliberative planning uses a world model to generate an optimal sequence of collisionfree actions
that can achieve a globally specified goal in a complex static environment.
However, in a dynamic environment, unforeseen obstacles may obstruct the action sequence,
and replanning to react to these situations can be too computationally expensive.
On the other hand, reactive control directly couples sensed data to appropriate actions.
It allows the robot to respond robustly and timely to unexpected obstacles and environmental changes
but may be trapped by them.
PROPOSED METHODOLOGY
The problem of goaldirected, collisionfree motion in a complex, unpredictable environment can be solved
by tightly integrating highlevel deliberative planning with lowlevel reactive control.
This work presents two such architectures for a nonholonomic mobile robot.
To achieve realtime performance, reactive control capabilities have to be fully realized so that
the deliberative planner can be simplified.
These architectures are enriched with reactive target reaching and obstacle avoidance modules.
Their target reaching modules use indirectmapping Extended Kohonen Map to provide finer and smoother motion control
than directmapping methods.
While one architecture fuses these modules indirectly via command fusion,
the other one couples them directly using cooperative Extended Kohonen Maps,
enabling the robot to negotiate unforeseen concave obstacles.
The planner for both architectures use a slippery cells technique to
decompose the free workspace into fewer cells, thus reducing search time.
Any two points in the cell can still be traversed by reactive motion.
PUBLICATIONS
 Enhancing the Reactive Capabilities of Integrated Planning and Control with Cooperative Extended Kohonen Maps.
Kian Hsiang Low, Wee Kheng Leow & Marcelo H. Ang, Jr.
In Proceedings of the
IEEE International Conference on Robotics and Automation (ICRA'03), pages 34283433, Taipei, Taiwan, May 1217, 2003.
Abstract. Despite the many significant advances made in robot motion research, few works have focused on the tight integration of highlevel deliberative planning with reactive control at the lowest level. In particular, the realtime performance of existing integrated planning and control architectures is still not optimal because the reactive control capabilities have not been fully realized. This paper aims to enhance the lowlevel reactive capabilities of integrated planning and control with Cooperative Extended Kohonen Maps for handling complex, unpredictable environments so that the workload of the highlevel planner can be consequently eased. The enhancements include fine, smooth motion control, execution of more complex motion tasks such as overcoming unforeseen concave obstacles and traversing between closely spaced obstacles, and asynchronous execution of behaviors.
 A Hybrid Mobile Robot Architecture with Integrated Planning and Control.
Kian Hsiang Low, Wee Kheng Leow & Marcelo H. Ang, Jr.
In Proceedings of the
1st International Joint Conference on Autonomous Agents and MultiAgent Systems (AAMAS02), pages 219226, Bologna, Italy, Jul 1519, 2002.
26% acceptance rate
Abstract. Research in the planning and control of mobile robots has received much attention in the past two decades. Two basic approaches have emerged from these research efforts: deliberative vs. reactive. These two approaches can be distinguished by their different usage of sensed data and global knowledge, speed of response, reasoning capability, and complexity of computation. Their strengths are complementary and their weaknesses can be mitigated by combining the two approaches in a hybrid architecture. This paper describes a method for goaldirected, collisionfree navigation in unpredictable environments that employs a behaviorbased hybrid architecture with asynchronously operating behavioral modules. It differs from existing hybrid architectures in two important ways: (1) the planning module produces a sequence of checkpoints instead of a conventional complete path, and (2) in addition to obstacle avoidance, the reactive module also performs target reaching under the control of a selforganizing neural network. The neural network is trained to perform fine, smooth motor control that moves the robot through the checkpoints. These two aspects facilitate a tight integration between highlevel planning and lowlevel control, which permits realtime performance and easy path modification even when the robot is en route to the goal position.
 Integrated Planning and Control of Mobile Robot with SelfOrganizing Neural Network.
Kian Hsiang Low, Wee Kheng Leow & Marcelo H. Ang, Jr.
In Proceedings of the
IEEE International Conference on Robotics and Automation (ICRA'02), pages 38703875, Washington, DC, May 1115, 2002.
Abstract. Despite the many significant advances made in robotics research, few works have focused on the tight integration of task planning and motion control. Most integration works involve the task planner providing discrete commands to the lowlevel controller, which performs kinematics and control computations to command the motor and joint actuators. This paper presents a framework of the integrated planning and control for mobile robot navigation. Unlike existing integrated approaches, it produces a sequence of checkpoints instead of a complete path at the planning level. At the motion control level, a neural network is trained to perform motor control that moves the robot from one checkpoint to the next. This method allows for a tight integration between highlevel planning and lowlevel control, which permits realtime performance and easy modification of motion path while the robot is enroute to the goal position.
 Integrated Robot Planning and Control with Extended Kohonen Maps.
Kian Hsiang Low.
Master's Thesis, Department of Computer Science, School of Computing, National University of Singapore, Jul 2002.
Singapore Computer Society Prize for best M.Sc. Thesis 20022003
Abstract. The problem of goaldirected, collisionfree motion in a complex, unpredictable environment can be solved by tightly integrating highlevel deliberative planning with lowlevel reactive control. This thesis presents two such architectures for a nonholonomic mobile robot. To achieve realtime performance, reactive control capabilities have to be fully realized so that the deliberative planner can be simplified. These architectures are enriched with reactive target reaching and obstacle avoidance modules. Their target reaching modules use indirectmapping Extended Kohonen Map to provide finer and smoother motion control than directmapping methods. While one architecture fuses these modules indirectly via command fusion, the other one couples them directly using cooperative Extended Kohonen Maps, enabling the robot to negotiate unforeseen concave obstacles. The planner for both architectures use a slippery cells technique to decompose the free workspace into fewer cells, thus reducing search time. Any two points in the cell can still be traversed by reactive motion.
VIDEO DEMOS
 Robot motion in an environment with unforeseen stationary obstacle
using command fusion
 Robot motion in an environment with unforeseen moving obstacle
using command fusion
 Robot motion in an environment that changes using command
fusion
 Robot motion in an environment with
unforeseen stationary concave and narrowly spaced convex obstacles using cooperative Extended Kohonen Maps
 Robot motion in an environment with
unforeseen moving obstacles using cooperative Extended Kohonen Maps
