Adjusting Weights in Artificial Neural Networks us

Authors: C Cotta E Alba R Sagarna P Larrañaga

Publish Date: 2002

Volume: , Issue: , Pages: 361-377

PDF Link

Abstract

Training artificial neural networks is a complex task of great practical importance Besides classical adhoc algorithms such as backpropagation this task can be approached by using Evolutionary Computation a highly configurable and effective optimization paradigm This chapter provides a brief overview of these techniques and shows how they can be readily applied to the resolution of this problem Three popular variants of Evolutionary Algorithms —Genetic Algorithms Evolution Strategies and Estimation of Distribution Algorithms— are described and compared This comparison is done on the basis of a benchmark comprising several standard classification problems of interest for neural networks The experimental results confirm the general appropriateness of Evolutionary Computation for this problem Evolution Strategies seem particularly proficient techniques in this optimization domain and Estimation of Distribution Algorithms are also a competitive approach

Keywords: