Neuroevolution of augmented topologies with difference-based mutation

Abstract

This study proposes the modification of the neuroevolution of augmented topologies, namely the difference-based mutation operator. The difference-based mutation changes the weights of the neural network by combining the weights of several other networks at the position of the connections having same innovation numbers. The implemented neuroevolution algorithm allows backward connections and loops in the topology, and uses several mutation operators, including connections deletion. The algorithm is tested on a set of classification problems and a rotary inverted pendulum problem and compared to the same approach without difference-based mutation. The experimental results show that the proposed weight tuning scheme allows significant improvements of classification quality in several cases and finding better control algorithms.