Comparison of EM algorithm and particle swarm optimisation for local model network training

Abstract

Local model networks (LMNs) offer a versatile structure for the identification of nonlinear static and dynamic systems. In this paper an algorithm for the construction of a tree-structured LMN with axis-oblique partitioning using particle swarm optimisation (PSO) is presented. The PSO algorithm allows the optimisation of arbitrary performance criteria but is only used for a certain subtask which helps to reduce the search space for the evolutionary algorithm very effectively. A comparison using an Expectation-Maximisation (EM) algorithm is presented. The differences and advantages of the LMN with PSO and the EM algorithm, respectively, are highlighted by means of an illustrative example. The practical applicability of the proposed LMN with particle swarm optimisation is demonstrated using real measurement data of an internal combustion engine.