Migration-based multiobjective genetic programming for nonlinear system identification

Nonlinear system identification is addressed by means of genetic programming. For a flexible selection of model structure and parameters, a multiobjective optimization of the tree encoded individuals is carried out, in terms of accuracy and parsimony. The paper suggests a new optimization algorithm based on the evolvement of two quasi-independent subpopulations, which makes use of a flexible migration scheme with adaptive thresholds and multiple rates. By efficiently exploiting the concept of dominance analysis, the algorithm is able to select compact and accurate models, with good generalization capabilities. The approach is compliant with nonlinear models, linear in parameters. That permits the hybridization with QR decomposition and the use of enhanced genetic operators, aimed to increase the algorithm convergence speed. The performances of the suggested design procedure are illustrated by the identification of two nonlinear industrial subsystems.