Toward an estimation of distribution algorithm for the evolution of artificial neural networks
Graham Holker, Marcus Vinicius dos Santos · 2010
This paper presents the preliminary results of a unique method of neuroevolution called Probabilistic Developmental Neuroevolution (PDNE). PDNE builds upon Gene Expression Programming (GEP) and Probabilistic Incremental Program Evolution (PIPE). Instead of building a Probabilistic Prototype Tree, as in PIPE, a Probabilistic Prototype Chromosome is built. The chromosome has a similar structure to a GEP chromosome (head, tail, and weight domain) and contains probabilities for each element of the gene. With this methodology, neural networks can be expressed in a similar manner to GEP, and solutions can be evolved via an Estimation of Distribution Algorithm. Preliminary results show promise, but further work is required to match the results of GEP.