Blocky Net: A New NeuroEvolution Method

Abstract

A long-term goal in evolutionary machine learning is to evolve compact models, which can learn effectively in complex, high dimensional domains with many noisy and irrelevant features. NeuroEvolution, the process of using evolutionary computation to evolve the topology and weights of neural networks is an effective and general approach to this. To date there has been little work using methods such as NeuroEvolution of Augmenting Topologies (NEAT), on very large, noisy datasets with up to 100,000 variables and there have been few alternative methods proposed. This paper proposes a new network called Blocky Net with built-in feature selection, and a limited maximum parameter space and complexity. This method is compared against NEAT and FS-NEAT in a supervised classification setting, using 20 diverse datasets ranging from 10 up to 100,000 features. The results show the proposed method is a valuable alternative to NEAT and FS-NEAT with better performance on 13 of the 20 datasets tested versus 2 for FS-NEAT, and is better than NEAT in all cases. The analysis also shows this minimal method of NeuroEvolution allows for a number of promising future research directions.