Some experimental evidence on the performance of GA-designed neural networks

While some efforts have been made to formalize the choice of parameters for artificial neural networks (multi-level perceptrons) using the backpropagation (BP) algorithm, studies reported in the literature to date appear to be dominated by use of heuristic methods of design. The objective of this paper is to test the comparative performance of neural networks with parameters determined by genetic algorithms (GAs). We have attempted to improve the quality of this type of research by comparing several models across seven real problem domains, including both regression and classification problems. Performance is benchmarked against neural networks whose parameters are selected using heuristic methods, as well as several widely used statistical models. For the problem domains tested, GA-assisted design yields models that consistently perform as well or better than heuristic methods and the statistical models.