Swarm intelligence based partitioning in local linear models identification

Abstract

A new evolutionary solution to the partitioning problem in local linear models (LLM) identification is developed. It consists in a master search process involving swarm intelligence (SI) based learning metaphor which trains the underlying system working space (SWS) oblique partitioning parameters, and a nested local optimization algorithm that estimates the LLM parameters in consequence. Finally two sequential outer incremental loops are used to select the LLM order and the optimal LLM network size respectively. The main advantages of this LLM identification approach are twofold: it is intended for simulation and prediction and is robust with respect to the LMM and the membership function (MSF) types. The effectiveness of the developed identification algorithm is confirmed by simulation.