Study of estimation of distribution algorithms applied to neuroevolution

This thesis proposes a methodology for the automatic design of neural networks via Estimation ofDistribution Algorithms (EDA). The method evolves both topology and weights. To do so, topol-ogy is represented with a fixed-length, indirect encoding; weights are represented as a bitwise en-coding. The topology and weights are searched via an incremental learning algorithm and a GuidedMutation operator. To explore suitable EDA ensembles, the study presented here interchangeablycombined two representations for topology, two for weights, and two learning algorithms. Testsused in the analysis include: XOR, 6-bit Multiplexer, Pole-Balancing, and the Retina Problem. Theresults demonstrate that: (1) the Guided Mutation operator accelerates optimization on problemswith a fixed fitness function; (2) the EDA approach introduced here is competitive with similarGP methods and is a viable method for Neuroevolution; (3) our methodology scales well to harderproblems and automatically discovers modularity.