Elitist multiobjective nonlinear systems identification with insular evolution and diversity preservation

The paper suggests a customized elite based genetic programming technique for the identification of complex nonlinear systems. The models generated by the proposed method are nonlinear, linear in parameters, as the universal approximation capacities of such a mathematical formalism have been rigorously proven. To better exploit the models´ parameter wise linearity, the authors propose a memetic approach that combines the stochastic structural transitions caused by enhanced genetic operators, with a deterministic parameter computation routine based on QR decomposition. This symbiosis assures a quasi simultaneous model structure and parameters selection and heightened search space exploration capabilities. To better fit the requirements of systems identification, the problem is formulated as a multiobjective optimization one, employing accuracy and parsimony assessment criteria. Two elitist evolutionary procedures have been implemented to obtain a solution, each featuring original contributions: the first one employs a dynamic clustering mechanism aimed at encouraging specific solutions of interest for the problem at hand, whilst the second is oriented towards maintaining population diversity by means of similarity analysis. The practical efficiency of the described methods is demonstrated relative to a multivariable test system with delayed inputs and a complex industrial plant.