Model Management for Decision Support
Last updated: July 4, 1996.An Overview of Model Management
Decision support systems (DSS) are computer-based systems that facilitate the use of data, models, and structured decision processes in decision making. Contributing disciplines---and useful keywords for search---include Decision Theory, Decision Analysis, Operations Research, Management Science, and Artificial Intelligence. The Decision Support Systems journal is a leading publication in this area. In this section, we aim to provide you with an overview to the World of Model Management.
Evolution of Model Management System
This evolution began in 1975 with the suggestion that decision models,
like stored data, are an important organizational information resource
that should be managed effectively and that specialized information
systems --- that is, model management systems --- should be developed for
this purpose. The purpose of a model management systems is to insulate
the users of a DSS from the physical aspects of model base storage and
processing, just as the purpose of a data management system is to
insulate users from the physical aspects of database storage and
processing. This suggests a duality between stored data and decisions
models, and the purpose of model management research is to extend our
existing knowledge of data management by investigating the properties of
systems in which the information objects of interest are not files but
rather are algorithms used to support decision processes. Research projects in
DISCS: are some of the modeling researches in our department.
What are models? Models to decision makers are instruments
which transform data into useful information. A model is an object or a
concept used to represent a real situation or an actual (physical)
machinery, system, etc. It is presented in a form that is scaled down
(physical model) or in an abstract framework that is well understood. A
model is a plan for information processing and provides a specification
for transforming information. Thus a model may be specified in
mathematical expressions, in natural language statements or in computer
programs. A mathematical model is an abstract (symbolic/algebraic)
representation which is made up of mathematical concepts involving
constants, variables, functions relationships, and restrictions.
Models manipulate input and stored data to yield results or output.
Such models update files, provide responses to user queries, and serve as
"black boxes" for performing such recurring analytical operations as
description, explanation, prediction, and resource allocation. Models are
linked to families of methodologies and range from the very simple to the
inordinately complex and span the gamut from operations research and
decision analysis to artificial intelligence and other sets of tool-boxes.
Model is a representation of some aspects of reality.
Problem is a description of something to be done with a
situation.
Solver is a manipulator of a model according to some
definite procedure for solving a problem or performing a task.
Model Instance is an instantiation of a mathematical model
type with its data. As data changes over time, instances of the same
model type will not necessarily be identical.
Model Base is a collection of model types contained in an
electronic storage medium and accessible to users and programs.
Modeling Approach refers to the discipline or field from
which concepts for designing a model management system have been
borrowed. The graph-based approach for model management, for example,
relies on concepts from graph theory for representing a mathematical
model.
Modeling Framework is a specialization within a modeling
approach. For example, Structured
Modeling is a framework within the graph-based approach. A purely
conceptual representation of a problem is built within a framework.
Modeling Language is a formal computer executable
notation which can be used to express the abstract concepts of a
framework. SML for instance, is a
modeling language developed for Structured Modeling Framework.
In the discussion of providing computer support for decision making,
many classifications of decision problems from different perspectives
have been used. In Alter's taxonomy of DSS, the classification focused on
the generic operations performed via the use of the DSS, not the type of
problem, functional area, decision perspective, etc. He mentioned:
Anothony's classification is based on functional areas in
strategic planning, tactical planning, operations control and industrial
control.
Simon simply classifies decision problems into two types:
programmed or structured decision, unprogrammed or
unstructured decision.
From Raiffa's decision theory, problems are classified
according to: single or multiple (group) decision makers; time staged
problem or otherwise; deterministic or uncertain; or single objective or
multiple objective. It is also been pointed out that recurrences of a
decision problem add its chances of getting computer support. Others also
argued the relevance of the style of decision makers. According to Raifa,
models are used to represent these real world decision problems fall into
three major categories:
A Modeling System is a computer system that accepts a
user requirement model, translate it into a form acceptable to a solver,
invokes the solver, and translates the output of the solver back into a
form that can be interpreted by the modeler. Solvers are not part of the
modeling system itself: they are merely utility systems that the modeling
system communicates with, usually developed by some independent party.
The modeling system must therefore communicate in two directions: with
the modeler and with the solvers. A good modeling system must make both
communications easy, i.e. the interface to the modeler must be oriented
towards the needs of a modeler, and the interface to the solvers must be
oriented towards the needs of the solvers.
A typical MS/OR modeling process involves a modeler's understanding of
the user's requirement model, selecting a solution method, translating
the requirement model to the solution design model, selecting a solver,
translating the solution design model to the input format of the solver,
invoking the solver and then interpreting the numerical results of the
solution.
Numerous modeling
languages such as GAMS and AMPL are now commercially available.
Model building involves the formulation of solution design
models into specific solver models and reformulation of solver models.
Model reformulation is required because of two reasons.
A modeling system supports model building. Existing modeling systems,
however, relies on modeler's expertise to transform the user requirement
model into solution design model that can eventually be translated by the
system into an executable solver model. Tools or knowledge-based systems
have been proposed to facilitate the model building task of transforming
a user requirement model into some representation(s) of the solution
design model. Two main approaches have been followed to development of
these tools or knowledge-based systems:Domain-specific and
domain -independent.
Prior to the development of modeling languages, models for a given
application were created from scratch and matrix generators were
developed to interface these models with solvers. The matrix generators
were generally written for a specific application and were not
immediately adaptable to a slightly different application environment.
This fact, combined with the programming skills required to develop these
generators, rendered this modeling approach unpopular among decision
makers. The development of algebraic
modeling languages such as GAMS and AMPL provided remedy for some of the
problems associated with model representation and execution. Other
approaches to model representation and creation includes:
Database approach. Advocates of the database approach
envision models being organized using a particular data model to insulate
users from the physical details of model base organization. Towards this
end, attempts have been made to represent mathematical models using the
CODASYL DBTG network data model, the entity-relationship model, and the
relational model.
Graph-based approach. In this approach, a mathematical
model is represented by one or more graphs or digraphs. The use of graphs
for knowledge representation has many advantages including conceptual
clarity, ease of programming and ease of manipulation. A graph consists
of nodes and arcs that capture the semantics of a model. Usually digraphs
are used where the nodes are interpreted as objects. Representation of
complex mathematical relationships through graphs results in a more
effective communication between the analysts and decision makers.
A recent and widely cited modeling framework that uses graphs as part
of its representational repertoire is the structured modeling.
A structured model representation of a problem is an organized,
partitioned and acyclic graph representing all the components of a
problem and the relationships between them.
Knowledge-based approach. In the knowledge
based-approach, Artificial
Intelligence(AI) tools and techniques
are applied to model management. Through various knowledge schemes the
syntactic knowledge of problem structures, the semantic knowledge of the
different components of a problem, and the procedural knowledge of how to
manipulate models can be represented. A variety of knowledge
representation schemes such as semantic networks, first order predicate
calculus, and production rules have been used for model representation
and management.
The model representation frameworks for knowledge-based model
management systems are 1)semantic nets and frame systems,
2)first order predicate calculus, and 3)production rules.
Model manipulation activities include model selection, model
scheduling, model integration, model invocation and new model components
generation.
This phase consists of checking the model assumptions, performing
sensitivity
analysis, and revising the model if necessary. The existing
model management systems provide very limited support for checking
assumptions and for model revisions. The capability to perform
sensitivity analysis can be found in GAMS, AMPL, GXMP, FW/SM, AIMM, and
PLATFORM.
Many of the currently available solvers provide sensitivity results
that cannot be easily understood by non-experts in modeling. A model
management system must provide user friendly screens that can project
these results in a manner that can be understood by non-experts. This can
be done through user interfaces which can display sensitivity results as
graphs or other easy-to-understand representations. A user-friendly
system should allow for user queries on the model solution or the model
itself by providing immediate expression evaluation as suggested by
Geoffrion.
An example of a system which is designed to assist analysts with their
use of linear programming models is ANALYZE.
This system not only
assists users in sensitivity analysis but also in model documentation,
verification, debugging, and result interpretation.
Contents:
Introduction
Model management is concerned with the representation and manipulation of
models and a model management system aims to provide (automated) support
for various phases of the decision modeling life cycle. Model management
is a software component in a decision support system.
Definitions of terms used commonly in Model
Management
Classification of Decision Problems
File-drawer systems, which
provide access to data items,
Data Analysis systems, which
allow the manipulation of data,
Analysis information systems,
which provide
the user with access to a series of databases and relatively small models,
Accounting models,
which calculate
the consequences of planned actions based on accounting definitions,
Representational models, which
generate
estimates of the consequences of actions on the basis of models which are
at least partially nondefinitional,
Optimization models, which
provide
guidelines for action by identifying optimal solutions consistent with
a set of constraints, and
Suggestion models,
which
yield a specific suggested decision for a relatively structured task.
Descriptive models are
defined
by a set of mathematical relations which simply predicts how a physical,
industrial or a social system may behave.
Normative models
constitute the
basis for decision making by a superhuman following an entirely rational
set of arguments. Hence quantitative decision problems and idealised
decision makers are postulated in order to define these models.
Prescribed
models involve systematic
analysis of problems as carried out by normally intelligent persons who
apply intuition and judgement.
Major functions of Model Management Systems
Creates models easily and
quickly,
either from scratch or from existing models or from building blocks.
Allows users to manipulate
the models so that they can conduct experiments and sensitivity
analysis ranging from "what-if" to "goal seeking".
Stores and manages a wide
variety of different types of models in a logical and integrated manner.
Accesses and integrates the model building blocks.
Catalogs and displays the
directory of models for use by several individuals in the organization.
Tracks models, data,
and application usage.
Interrelates models with
appropriate linkages through the database.
Manages and maintains
the model base with
management functions analogous to database management: store, access,
run, update, link, catalog, and query.
Modeling Systems
Model Building
Conforming to the structure
required in
the specific solver is necessary for using the solution algorithms.
Solver model could be
too complex
to be handled by any solver, or better efficiency of using the solution
algorithms is desired. To reduce the complexity, two approaches can be
taken: simplifying the model composition, for example, by removing
redundancies in the solver model, and decomposing the structure of the
solver model.
Approaches to Model Representation and Creation
Model Manipulation
Model Selection In this
function, a
support system tries to figure out what models are available to be used
for the solution design model and then automatically selects or allows
the user to select a model for execution. To the user, the emphasis of
the selection process is on what the model type is capable of doing and
not on how the solution procedure works. Usually, different types of
information will be needed to select a model type and it is impossible to
deal with all such information at the same time. Thus there is need for a
stepwise refinement process of the model selection procedure that
navigates through the information in a systematic manner.
Model Scheduling. This
function
involves arranging a set of models into an execution sequence so that the
result of one model is used by the next one. Model scheduling leads to a
sequence of models in their execution orders based mainly on input/output
relationships. That is, if the outputs of model A are inputs of model B,
model A is executed before model B and the result of model A is delivered
to model B.
Model Integration.
This function consists of identifying and properly combining models
and other DSS components (such as data files and data analysis procedures)
needed to respond to a specific query. This may require the use of intelligent
software, since integration may be viewed as inferring a query schema
from a set of information schemas. Intelligent systems have been
developed for this purpose based on logic programming, connection graphs,
AND/OR graphs, heuristic search, rules, semantic nets, frames. Some of
these ideas have been extended to the integration of distributed model bases.
Model
Invocation.This function
involves activation of the selected or integrated models to solve the
specific modeling process initiated by the user's problem.
New model components
generation. It
is possible that no existing model is capable of solving the problem.
This might arise under two circumstances. The basic model components
already reside in the model base, but required functions or tests are not
formed. Otherwise, neither the basic model of the new functions or tests
with existing are in the model base. The former requires the generation
of a new model from scratch. Both solutions are entitled as the process
of "new model components generation" in the logical flow. Note that model
integration does not generate new formulas; it merely reorganizes model
schemas on hand.
Result Interpretation
DSS Assets currently available on the Web
Here's a list of URLs that indicate WWW locations having strong relevance to DSS. I am not always able to keep this updated
or even to continually check for relevance. So any comments for
additions and deletions would be most
appreciated. You can send email (preferably---if you suggest new entries---with new entries set up in HTML-LI format)
to yeogk@iscs.nus.sg
With each URL is a brief description: ``Quoted sentences'' were pulled off from the original sources;
Other statements represent my remarks.
General Online Information on Decision Support Technologies and
Methods
General Modeling Systems
Online Databases useful in DSS Work
Netlib is ``a repository of mathematical software, data, documents, address lists, and other useful
items.''
WWW Links to Related Fields
e-mail wonglips@iscs.nus.sg