Title

Knowledge sharing in multiagent environments: Neural networks as intelligent agents

Date of Award

1997

Availability

Article

Degree Name

Doctor of Philosophy (Ph.D.)

Department

Electrical and Computer Engineering

First Committee Member

Mansur R. Kabuka, Committee Chair

Abstract

Neural networks are examined as candidates for Intelligent Agency. Intelligent Agent architectures, applications and definitions are surveyed, and a working definition of Agents is derived. A solution to neural network knowledge sharing and communication problems in multiagent environments is proposed. Reductions in translation complexity and the inherent value of standardized Interlinguas for knowledge sharing in multiagent systems are examined. The proposed ANSI X3T2 standard, Knowledge Interchange Format (KIF), is selected as the Interlingua of choice for communication and sharing knowledge between neural network agents.This thesis introduces ABTRE (Absolute Binary Tree Rule Extraction), an efficient algorithm for extracting propositional, decompositional symbolic rules from neural networks with common architectures. ABTRE imposes no heuristic limitations on the number of extracted rules or the number of antecedents per rule, but is capable of doing so. ABTRE reduces rule extraction computation times via automatic elimination of rule redundancies and subsumptions, and automatically supports negated antecedents.The thesis also introduces LPSRI (Linear Programming Symbolic Rule Insertion), a methodology utilizing two algorithms and Linear Programming (LP) to insert symbolic rule sets into target neural networks. LPSRI transforms symbolic rule sets into the virtual node weights and biases required to initialize target neural networks. Two algorithms create sets of "direct" and "indirect" LP constraints from the symbolic rule set, which restrain the resulting solution from inducing target network behavior not inherent in the original source network.The accuracy of the ABTRE and LPSRI algorithms are verified by empirical evidence and experimental results. Verification on a variety of noisy, publicly available datasets indicates that the combined algorithms are capable of extracting and reinserting symbolic rule sets that reproduce the behavior of source networks.

Keywords

Information Science; Artificial Intelligence; Computer Science

Link to Full Text

http://access.library.miami.edu/login?url=http://gateway.proquest.com/openurl?url_ver=Z39.88-2004&rft_val_fmt=info:ofi/fmt:kev:mtx:dissertation&res_dat=xri:pqdiss&rft_dat=xri:pqdiss:9824521