Genetic Programming for System Identification

Abstract

This chapter discusses the features of genetic programming based identification approaches, starting with the connected theoretical background. The presentation reveals both advantages and limitations of the methodology and offers several recommendations useful for making GP techniques a valuable alternative for mathematical models’ construction. For a sound illustration of the discussed design scheme, two GP-based multiobjective algorithms are suggested. They permit a flexible selection of nonlinear models, linear in parameters, by advantageously exploiting their particular structure, thus improving the exploration capabilities of GP and the interpretability of the resulted mathematical description. Both model accuracy and parsimony are addressed, by means of non-elitist and elitist Pareto techniques, aimed at adapting the priority of each involved objective. The algorithms’ performances are illustrated on two applications of different complexity levels, namely the identification of a simulated system, and the identification of an industrial plant.