Parallel Soft Computing Models and their Applications

Abstract

This thesis presents sound mathematical modelling and extensive empirical performance newlinestudies on interval type-2 fuzzy sets based neural networks and evolutionary learning newlinestrategies along with their parallel implementations. Further, three novel models of fuzzy newlinesystems have been proposed viz., (i) Interval Type-2 Subsethood Neural Fuzzy Inference newlineSystem (IT2SuNFIS) (ii) Interval Type-2 Mutual Subsethood Fuzzy Neural Inference System- newline1 (IT2MSFuNIS-1) and (iii) Interval Type-2 Mutual Subsethood Fuzzy Neural Inference newlineSystem-2 (IT2MSFuNIS-2). It has been shown that three models compare favourably or newlineoutperform existing models in eight different benchmark problems. newlineIn the proposed models, which embed the interval type-2 fuzzy knowledge using newlineMamdani if-then rules, the signal transmission through the input-rule layer for IT2SuNFIS takes place using subsethood composition; in IT2MSFuNIS-1 and IT2MSFuNIS-2, a mutual subsethood measure has been used. Expressions for interval type-2 mutual subsethood newlinemeasure have been derived and it is shown empirically that mutual subsethood is better newlinethan subsethood measure. The universal function approximation property of the proposed models has been proved using Stone-Weierstrass theorem. newlineDifferent strategies are used for training of models, such as Artificial Bee Colony newlineDifferential Evolution, memetic procedures employing DE and gradient descent and newlinedifferential evolution. These are realized and investigated on a parallel platform using newlineLAM/MPI (Local Area Multicomputer/ Message Passing Interface). The simulation results newlinefor the eight application areas are summarized below: newlineIn the case of benchmark function approximation problem given by Narazaki and newlineRalescu, the function approximation capabilities of the proposed IT2SuNFIS and newlineIT2MSFuNIS-1 models are found to be superior to existing models reported in the literature. newlineFor a time-series prediction benchmark problem, viz., Mackey-Glass, IT2SuNFIS compares favourably to similar kind of models with lesser number of free parameters;