Increasing crossover operator efficiency in multiobjective nonlinear systems identification

An elitist multiobjective optimization methodology, based on genetic programming, is suggested in the following, as means of identifying complex nonlinear systems. The structure and parameters of the nonlinear models are selected simultaneously as result of the conjoint usage of customized genetic operators and of a deterministic parameter computation procedure. This symbiosis is configured to efficiently exploit the nonlinear, linear in parameters formalism, a proven universal approximator, according to which the models are generated. In order to protect useful model terms from fragmentation via crossover, the authors have introduced a novel encapsulation mechanism supervised by a fuzzy controller. To meet the specific requirements of systems identification in engineering applications, the optimization procedure considers two evaluation criteria, namely accuracy and parsimony, exploited from an elitist standpoint. The approach also features an original similarity analysis technique, meant to encourage population diversity. The practical efficiency of the proposed identification algorithm was tested in the framework of a real life industrial system.