Evolving neural networks in environments with delayed rewards by a real-coded GA using the unimodal normal distribution crossover

The Neuro-Evolution (NE), the training of neural networks with genetic algorithms (GAs), has received much attention as one of the reinforcement learning techniques that can let agents learn appropriate policies, i.e. mappings from sensory inputs to action outputs, in environments with delayed rewards. Although several studies on NE systems have been made so far, there are no studies that take account of epistasis among weight parameters in training neural networks with GAs. In function optimization, epistasis among parameters is one of the important features which make functions difficult to be optimized. Epistasis among parameters has to be considered in order to successfully optimize difficult functions with large number of parameters to be determined. In this paper, we present an NE system based on a real-coded GA using the Unimodal Normal Distribution Crossover (UNDX). The UNDX shows excellent performance in optimizing functions with strong epistasis among parameters. The NE system based on the UNDX are applied to some benchmark problems, which are more difficult than those used in previous work. The results suggest that epistasis among weight parameters should be considered when we train neural networks for difficult tasks by NE systems.