V-MDF
(Visualizer for Metaheuristics Development Framework)

This work has been presented in Metaheuristics International Conference 2005 @ Vienna, Austria (click here to read the conference paper) and in the post-conference volume (Metaheuristics - Progress in Complex Systems Optimization).

Overview

The main concepts/issues raised in this line of work are as follows:

1. New classification of metaheuristics tuning problem.

Metaheuristics tuning problem (finding the best configuration for a metaheuristic to solve the underlying combinatorial optimization problem) is actually quite broad, this term `tuning' is `abused'. Therefore, we proposed a new classification of tuning problem by dividing it into three types. These three types are:

a. type-1: calibrating parameter values.
b. type-2: choosing best components.
c. type-3: tuning search strategies.

We argue that all three types must be properly addressed to solve tuning problem well (the type-3 is the one that is commonly overlooked by metaheuristic users). We support our case by quoting comments from various well-known researchers regarding this tuning problem and presented a simple case study where addressing type-1 and type-2 only, will not give us the best solution for tuning problem...

2. Review of recent works around tuning methods for addressing tuning problem

We observed an emerging trend of research around this topic (see the publication year of these works!). Some examples are: F-Race (Birattari, 2004), CALIBRA (Adenso-Diaz and Laguna, 2006), Visualization of Search Behavior (Kadluczka et al., 2004), various statistical methods: Fitness Distance Correlation, Run Time Distribution, etc (Hoos and Stuetzle, 2005), self-configuring/self-adaptive algorithms, etc...

3. Visual Diagnosis Tuning methodology

We proposed a new way to conduct tuning by combining the power from human (visualization, intelligence, etc) and machine (speed, endurance, etc). The rules in form of {cause-action-outcome} tuple is presented. This tuple is quite straightforward and commonly used to describe natural human behavior in psychology (see, action, observe the result)...

4. Visualizer for Metaheuristics Development Framework (V-MDF)

We built a tool to support visual diagnosis tuning methodology. This tool consists of two parts:

a. Distance Radar visualization:
    To economically display visualization of search trajectory (search space size is exponential) to user.
b. Rule-base:
    To help the algorithm designer to be more objective in choosing which search strategy to choose.

Hands-On Demo

To get a glimpse of how Visual Diagnosis Tuning works, please download the following items:

1. V-MDF executables:
   VMDF MTP solver (before tuning ~ 850kb)
   VMDF MTP solver (after tuning ~ 850kb)

2. The Military Transport Planning (MTP) test instances used in our paper,
   the test file format is included: (click here). The artificial test case generator is here.

3. OpenGL + CsGL library files (196kb). Extract the zip file into any folder and run the "libinstall.bat" to copy the required *.dll-s to your \windows\system32 folder. We used these graphic libraries to draw the visualization.

And then follow these steps:

1. After you have downloaded the files, run VMDF_MTP_Before_Tuning.exe

2. Type in the full path and the file name of an MTP test instance.

3. V-MDF will run...

4. Please observe what is being drawn in the distance radar + textual output in the background

5. Compare what you see with the results from VMDF_MTP_After_Tuning.exe...

The discussion on how to interpret V-MDF Distance Radar and Rule Base outputs, more explanation about the MTP problem itself, the reasoning on how our artificial test case generator works, various ways that we know to attack this MTP problem, etc, will be written soon...

Future Works

V-MDF has been succeeded by our next generation tool, which implements similar Visual Diagnosis Tuning concept, but in a more advanced way: Viz - SLS Engineering Suite (click here to go to Viz web page).