WPI Computer Science Department

Computer Science Department
------------------------------------------

Flexible Learning in Multi-Agent Systems


Researcher:

Dan L. Grecu (dgrecu@cs.wpi.edu)
Ph.D. research: completed May 2000
Advisor: Dave Brown
External: Diana Gordon (Naval Research Laboratory).

Abstract:

Intelligent design systems have reached a level of complexity where it becomes increasingly hard for developers to completely anticipate the variety of situations that may occur at design time. This can result in the failure of the design systems to produce a design solution when one exists, or in 'weaker' solutions than the ones that are actually available in a given problem context.

This work develops a learning-based approach for multi-agent design systems. It provides design agents with the ability to correct and to improve the design decision-making knowledge with which they initially set out to solve design problems. The proposed flexible learning paradigm allows agents to identify the situations in which their design decision-making fails or can be enhanced, and to respond by initiating learning processes.

Agent learning consists of the acquisition of expectations, a form of predictive knowledge that helps agents to avoid the regions of the design space with no solutions or with 'weak' design solutions. Learning proceeds by causal attribution and covariational analysis, using wrappers for inductive learning. Agents validate and monitor the quality of the expectations produced.

The thesis describes the role of expectations in design, and how expectations can be acquired, validated and used. It describes the LEAD multi-agent design system developed to investigate the proposed agent learning model. The experiments with LEAD demonstrate how developing expectations can improve multi-agent design solutions, and how expectation learning provides agents with the autonomy to adapt to the particular situations where they consider learning to be needed.

The conclusions drawn from the experimental work highlight the strengths of the expectation-based learning approach in multi-agent design, such as performance and quality improvements, and identify directions for further investigation.

Publications:


[Feedback] [Search Our Web] [Help & Index]

[Return to the WPI Homepage] [Return to the CS Homepage]

dcb@cs.wpi.edu / Fri May 5 16:18:13 EDT 2000