A NewFastAlgorithm forFuzzyRuleSelection

Thispaperinvestigates theselection offuzzy rules forfuzzy neural networks. Themainobjective istoeffectively andefficiently select therules andtooptimize theassociated parameters simultaneously. Thisisachieved bytheproposal of afast forward ruleselection algorithm (FRSA), wheretherules areselected onebyoneandaresidual matrix isrecursively up- dated incalculating thecontribution ofrules. Simulation results showthat, theproposed algorithm canachieve faster selection offuzzy rules incomparison withconventional orthogonal least squares algorithm, andbetter network performance thanthe widely usederror reduction ratio method(ERR). I.INTRODUCTION Fuzzyneural networks represent alarge class ofneural networks that combine theadvantages ofassociative memory networks (e.g. B-splines, radial basis functions andsupport vector machines) withimproved transparency, acritical issue fornonlinear modelling using conventional neural networks. Forassociative neural networks, theadvantage isthat thelin- earparameters canbetrained online withgoodconvergence andstability properties. However, theyproduce essentially black boxmodels withpoorinterpretability . Forfuzzy neural networks (FNNs), thebasis functions areassociated with somelinguistic rules, andeverynumerical result canadmit alinguistic interpretation (1).