Abstract

Multi-agent systems, systems consisting of more than one acting and decision making entities, are of great interest to researchers because they have advantages for some specific tasks where it would be more effective to use multiple small and simple robots rather than a large and complex one. One of the major problems with multi-agent systems is developing a means to organize or control the overall behavior of the system. Typically, multi-agent control involves one of two structures. In some designs, there is a hierarchy with some robots being leaders and other followers. Other designs involve robot specialization towards one particular task or individual robots which loosely or strongly cooperate in some manner to yield the desired behavior.

This thesis studies using bayesian decision networks (BDNs) as a method to control individual robots to achieve some group or cooperative behavior. BDNs are powerful tools enabling designers of intelligent agents to model the agent's environment and the behavior of other agents without expert knowledge about a system. The probabilistic nature of these networks allows agents to learn about themselves and their environment by updating their bayesian network (BN) with new observations. While two methods of learning and responding to change in the environment with BNs, parameter learning and structure learning, have been studied by many researchers as a means to control a single robot or teams of robots, a third method, utility updating, has seen little study. This work is thus a novel study of BN control since it incorporates all three methods to develop a decision theoretic agent (DTA).

The agent is applied to a modified version of a personal rapid transit (PRT) problem (or personal automated transport (PAT)) that is simulated in Matlab. PRT is a proposed public transport method which offers automated on-demand transportation between any two nodes of the transportation network. The PRT problem of interest is that of autonomous control. This can be likened to one of multi-agent control of many identical agents.

Several agents are developed to solve the problem, a rule based agent and BN-agents which use various subsets of the three network updating methods. The experimental results show that the DTA that uses parameter, structure, and utility updating could be a superior solution to agents based only on some subset of those methods.

Library of Congress Subject Headings

Bayesian statistical decision theory--Data processing; Intelligent agents (Computer software)

Publication Date

2007

Document Type

Thesis

Student Type

Graduate

Degree Name

Electrical Engineering (MS)

Department, Program, or Center

Electrical Engineering (KGCOE)

Advisor

Ferat Sahin

Advisor/Committee Member

Daniel Phillips

Advisor/Committee Member

Eric Peskin

Comments

Physical copy of thesis available from RIT's Wallace Library at QA279.5 .U64 2007

Campus

RIT – Main Campus

Share

COinS