Multiobjective Genetic Programming for Nonlinear System Identification

The paper presents a novel identification method, which makes use of genetic programming for concomitant flexible selection of models structure and parameters. The case of nonlinear models, linear in parameters is addressed. To increase the convergence speed, the proposed algorithm considers customized genetic operators and a local optimization procedure, based on QR decomposition, able to efficiently exploit the linearity of the model subject to its parameters. Both the model accuracy and parsimony are improved via a multiobjective optimization, considering different priority levels for the involved objectives. An enhanced Pareto loop is implemented, by means of a special fitness assignment technique and a migration mechanism, in order to evolve accurate and compact representations of dynamic nonlinear systems. The experimental results reveal the benefits of the proposed methodology within the framework of an industrial system identification.