A Learning Algorithm for Local Linear Neuro-fuzzy Models with Self-construction through Merge & Split

Abstract

A self-constructing version of locally linear model tree (LOLIMOT) algorithm for structure identification in neuro-fuzzy models is proposed in this paper. LOLIMOT is an incremental tree-construction learning algorithm that partitions the input space by axis-orthogonal splits. In each iteration, LOLIMOT splits a local model into two models in a way that a local classification error is minimized. As a result, during the training procedure some of the formerly made divisions may become suboptimal or even superfluous. In this paper, the LOLIMOT is improved in two ways: (1) the ability to merge previously divided local linear models is added, and (2) a simulated annealing stochastic decision process is responsible to select a local model for splitting. Comparing to the LOLIMOT, our proposed improved learning algorithm shows the ability to construct models with fewer number of rules at comparable modeling errors. Algorithms are compared through a case study of nonlinear function approximation. Obtained results demonstrate the better performance of modified method as compared to that of original form of the LOLIMOT algorithm