Session:

LBP - Late Breaking Papers

Title:

A Comparative Analysis of Simplification and Complexification in the Evolution of Neural Network Topologies

Authors:

Derek James
Philip Tucker

Abstract:

Approaches to evolving the architectures of artificial neural networks have involved incrementally adding topological features (complexification), removing features (simplification), or both. We will present a comparative study of these dynamics, focusing on the domains of XOR and Tic-Tac-Toe, using NEAT (NeuroEvolution of Augmenting Topologies) as the starting point. Experimental comparisons are done using complexification, simplification, and a blend of both. Analysis of the effects of each approach on the variation, complexity, and fitness of the evolving populations demonstrates that algorithms employing both complexification and simplification dynamics search more efficiently and produce more compact solutions.