Chapter

Genetic and Evolutionary Computation – GECCO 2004

Volume 3103 of the series Lecture Notes in Computer Science pp 69-81

Evolving Reusable Neural Modules

  • Joseph ReisingerAffiliated withDepartment of Computer Sciences, The University of Texas at Austin
  • , Kenneth O. StanleyAffiliated withDepartment of Computer Sciences, The University of Texas at Austin
  • , Risto MiikkulainenAffiliated withDepartment of Computer Sciences, The University of Texas at Austin

* Final gross prices may vary according to local VAT.

Get Access

Abstract

Topology and Weight Evolving Artificial Neural Networks (TWEANNs) have been shown to be powerful in nonlinear optimization tasks such as double pole-balancing. However, if the input, output, or network structures are high dimensional, the search space may be too large to search efficiently. If the symmetries inherent in many large domains were correctly identified and used to break the problem down into simpler sub-problems (to be solved in parallel), evolution could proceed more efficiently, spending less time solving the same sub-problems more than once. In this paper, a coevolutionary modular neuroevolution method, Modular NeuroEvolution of Augmenting Topologies (Modular NEAT), is developed that automatically performs this decomposition during evolution. By reusing neural substructures in a scalable board game domain, modular solution topologies arise, making evolutionary search more efficient.