Configuring the parameters of artificial neural networks using neuroevolution and automatic algorithm configuration

Abstract

NeuroEvolution (NE) is a powerful method that uses Evolutionary Algorithms (EAs) for learning Artificial Neural Networks (ANNs). However, NE's performance is determined by the definition of dozens of parameters that guide the search of the EAs. In this study we apply automatic algorithm configuration for tuning the parameters of a NE method in an offline matter. The tuned NE method is then used to evolve the weights, topology and activation functions of ANNs while performing feature selection and its performance is compared to the case of using default parameters. We show that tuning the parameters results in NE methods able to solve the problems with 100% accuracy in significantly less generations.