A regressive schema theory based tool for GP evolved nonlinear models

Nonlinear systems identification is approached by employing a genetic programming computational tool featuring explicit building block exploitation. The level of adaptation of recurrent model sub-structures is assessed by a fuzzy module. The first contribution of the paper resides in using the fuzzy classification results to reconfigure the cut point selection probabilities of regressor inner nodes, a process called encapsulation. This allows for the second innovation, namely the design of context aware genetic operators capable of protecting the existing instances of fit building blocks and of creating new ones. The computational costs of encapsulation are reduced by employing a novel regressive schema theory - the third and main paper contribution - which assesses the inherent chances of regressor survival. A thorough theoretical support for demonstrating the efficiency of context aware operators in transmitting schema instances over the generations is introduced. The suggested algorithm is experimentally validated in the framework of a complex, industrial, nonlinear subsystem of a sugar factory.