Evolutionary Artificial Neural Networks That Learn and Generalise Well

Evolutionary artiicial neural networks (EANNs) refer to a special class of artiicial neural networks (ANNs) in which evolution is another fundamental form of adaptation in addition to learning. The evolution in EANNs is often simulated by genetic algorithms (GAs), evolutionary programming (EP), or other evolutionary algorithms. This paper describes an EP-based EANNs which learn both their weights and architectures through the combination of a hybrid learning algorithm and the EP algorithm. A nonlinear ranking scheme and ve mutation operators are used in our EP. These ve mutation operators are applied sequentially and selectively to each individual in a population. Such sequential application encourages the evolution of smaller ANNs with fewer hidden nodes and connections. We have tested our EP-based EANNs on the parity problem of various sizes. Very good results have been achieved. For example, a three hidden node feed-forward ANN can be evolved for the 9-parity problem. In order to improve the generalisation capability of our EP-based EANNs, we have introduced two validation sets in the combined evolution and learning process. Such an approach can improve the generalisation ability of our EP-based EANNs signiicantly. Our experimental study with a credit card problem has connrmed that EP-based EANNs can generalise well, at least for the problem concerned. In fact, it produces one of the best results we are aware of.