Hybrid computational intelligence models with symbolic rule extraction for pattern classification
Loading...
Date
2008-12
Authors
Ali Quteishat, Anas Mohammad
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
This thesis is concerned with the development of hybrid Computational Intelligence
(CI) models for tackling pattern classification problems. The research focuses on
hybridization of the Artificial Neural Network (ANN) and Genetic Algorithm (GA)
models for data-based learning and rule extraction. In addition, a novel Multi-Agent
Classifier System (MACS) is devised to combine the hybrid Cl models into a
common framework for pattern classification.
The Fuzzy Min-Max (FMM) network is employed as the backbone for developing,
the hybrid Cl models in this research. Two novel modifications are proposed to
improve the performance of FMM and to extract symbolic if-then rules from its
knowledge base. The first modification aims to reduce the number of hyperboxes
formed in FMM and to enhance its performance. The second modification
incorporates the GA to extract fuzzy if-then rules from FMM-based networks.
To further improve the robustness and performance of FMM-based networks for
pattern classification, a new MACS, developed based on the Trust-Negotiation-
Communication (TNC) reasoning model of multi-agent systems, is introduced. A
novel trust measurement method based on the Bayesian decision theory is
formulated. The performances of the MACS are evaluated using benchmark as well
as real industrial and medical problems, and the bootstrap method is used to quantify
the performances statistically. The results ascertain the effectiveness of employing
the MACS model for undertaking pattern classification tasks, along with
comprehensible if-then rules for explaining their predictions to domain users.
Description
Keywords
Intelligence models with symbolic , Pattern classification