This paper introduces a novel, general-purpose neuroevolution method which is named Broadened neuroEvolutionary Algorithm for Searching Topologies (BEAST). The goal of this algorithm is to solve premature convergence in evolution in dynamic problems or environments. Premature convergence is one of the major causes of failure in evolutionary artificial neural networks (EANNs), since it slows or halts evolution when local optima or changes to the objective function are present. BEAST combines several diff erent techniques to reduce the e ffects of convergence, each of which targets one or more of its major causes. The most radical of these techniques applies a meta-search to direct evolution from a higher level than previous approaches. This thesis tests the performance of BEAST against convergence in three major types of dynamic problems.

PDF

Technical Report HR-11-02, Department of Computer Science, The University of Texas at Austin, 2011.

@techreport{dewet:ugthesis11, title={Avoiding Premature Convergence in NeuroEvolution by Broadening the Evolutionary Search}, author={Matthew de Wet}, number={HR-11-02}, institution={Department of Computer Science, The University of Texas at Austin}, type={Undergraduate Honors Thesis}, url="http://nn.cs.utexas.edu/?dewet:ugthesis11", year={2011} }