Information-dependent switching of identification criteria in a genetic programming system for system identification

Abstract

Genetic Programming (GP) can be used to identify the nonlinear differential equations of dynamical systems. If, however, the fitness function is chosen in a classical way, the optimization will not work very well. In this article, we explain the reasons for the failure of the GP approach and present a solution strategy for improving performance. Using more than one identification criterion (fitness function) and switching based on the information content of the data enable standard GP algorithms to find better solutions in shorter times. A computational example illustrates that identification criteria switching has a bigger influence on the results than the choice of the GP parameters has. © Springer-Verlag Berlin Heidelberg 2006.