Creation and optimization of fuzzy inference neural networks
Abstract
Fuzzy Inference Neural Networks (FINN) are fuzzy-rule based systems that allow the use of systematic algorithms to initialize and optimize them. The initialization consists, basically, of a clusterization algorithm, while for their optimization, gradient methods and genetic algorithms have been used. Unlike crisp neural nets, the resulting FINN lend themselves to interpretation. All these advantages are important, but they would not be very attractive if there is not a chance, at least in theory, of approximating the training set below a given accuracy. Part of this dissertation is dedicated to show a constructive proof of this property of the FINN. Although FINN have advantages over crisp neural nets and other types of fuzzy inference systems, their creation and optimization processes suffer from disadvantages. Among others, the initialization procedure is sensitive to the order of the examples in the training set; the optimization phase requires intense computation and the existing genetic algorithms do not take into consideration the existing relationships among their parameters. Another part of this document details a new genetic algorithm which alleviates the stated disadvantages. In addition, with the purpose of speed up, a parallel implementation of the algorithm is described. The results of using the parallel implementation of the algorithm for three machine learning benchmark problems are presented. Finally, some conclusions and new lines research are proposed.
Subject Area
Computer science|Electrical engineering
Recommended Citation
Cruz-Cano, Raul, "Creation and optimization of fuzzy inference neural networks" (2005). ETD Collection for University of Texas, El Paso. AAI3167939.
https://scholarworks.utep.edu/dissertations/AAI3167939