Neuroeboluzio Informazioaren Ustiapena: Iragana Aztertu Etorkizun Eraginkor Baterako

Abstract

Neuroevolutionary algorithms, using evolutionary techniques to optimize neural network structures, are computationally costly but widely applied due to their strong performance. Typically, the best-found structure is the only outcome of the algorithm, while valuable residual information from the search is overlooked. This paper proposes a method to extract and leverage this information to build a metamodel that enhances future neural architecture searches. By analyzing optimal generative adversarial network structures, a Bayesian network-based meta-model is proposed, which can improve search initialization, and overcome local optima in structural optimization.