Lowering Evolved Artificial Neural Network Overfitting through High-Probability Mutation

Abstract

Artificial Neural Networks often suffer from overfitting, both when trained through backpropagation or evolved through a Genetic Algorithm. An attempt at mitigating the overfitting of GA-evolved ANNs is made by using High-Probability Mutation (≈0.95) on binary-encoded ANN weights. The benchmark used is predicting the evolution of an Internet social network using real-world data. A lower bound is put on the overfit, and both prediction error and overfit are further broken down according to ANN hidden-layers size.