Optimal Construction of Local Model Networks Using Genetic Algorithm

ABSTRACT Local Model (LM) networks represent a complex nonlinear dynamical system by a weighted sum of locally valid, simpler sub-models defined over small regimes of the operating space. Constructing such networks requires the determination of the appropriate regimes and the local model parameters. This paper describes a new soft computing approach to simultaneously determine the local model parameters and the appropriate operating regimes. Structured genetic coding is used to obtain the optimal number of ARX local models and their parameters. Centres and covariances of Gaussian interpolation functions are found through genetic evolution with fuzzy logic used to dynamically evolve the genetic algorithm. Several modifications to the classical genetic algorithm are proposed to detemline the locally valid sub-models. The parallel model prediction error is minimized to produce good generalization from the identified LM network. The modelling potential of the resultant LM network is investigated on practical data from an experimental pH neutralisation process. The same data was also used in a recent comparative study of a hybrid training algorithm and tree type construction (McGinnity and Irwin, 1999) for LM networks. It is concluded that the LM network from the new GA based construction formed a significantly better non linear dynamical model of the pH process.