A Hybrid Genetic Algorithm and Radial Basis Function NEAT

Abstract

We propose a new neuroevolution technique that makes use of genetic algorithm to improve the task provided to a Radial Basis Function - NEAT algorithm. Normally, Radial Basis Function works best when the input- output mapping is smooth, that is, the dimensionality is high. However, if the input changes abruptly, for example, for fractured problems, efficient mapping cannot happen. Thus, the algorithm cannot solve such problems effectively. We make use of genetic algorithm to emulate the smoothing parameter in the Radial Basis function. In the proposed algorithm, the input- output mapping is done in a more efficient manner due to the ability of genetic algorithm to approximate almost any function. The technique has been successfully applied in the non- Markovian double pole balancing without velocity and the car racing strategy. It is shown that the proposed technique significantly outperforms classical neuroevolution techniques in both of the above benchmark problems. © Springer International Publishing Switzerland 2014.