This paper presents a Genetic Algorithm (GA) designed for finding topologies of recurrent networks in order to optimize the performance in terms of learning error and generalization ability. In this context, the fitness of a chromosome (network) is a function of its estimated test error, i.e., its estimated generalization ability. For this spesific GA, two different fitness-functions are tested, one based on a validation set, and one based on Akaikes's Final Prediction Error estimate (FPE).
In this study, the task of the Neural Net is to predict the sunspot activity based on the activity of preceeding years. We find that the first fitness-function does not give satisfactory evolution for this task. The FPE based gives better agreement between the estimated test error and the actual test error, end hence better performance of the GA. However, it needs corrections for different topologies. This is dicussed in the last section.