On A Genetic Algorithm for the Selection of Optimally Generalizing Neural Network Topologies

A genetic algorithm is used for the topology design of feed-forward neural networks. The generated neural network topology populations are then trained to approximate non-linear relationships of multiple variables. A specifically designed fitness function using the epistemological principle of dimensional homogeneity is used for the evaluation of the individual neural network generations and for the selection of the best generalizing neural networks topologies. A general theory for the topology design and the explanation of the correct generalization capability of non-linear feed-forward neural networks is developed, mathematically proved and explained using simulation results.