Towards Designing Arti cial Neural Networks by Evolution

Designing artiicial neural networks (ANNs) for different applications has been a key issue in the ANN eld. Although there are many training algorithms available for learning ANN's connection weights, algorithms which can learn ANN's architectures are relatively few. At present, ANN design still relies heavily on human experts who have suucient knowledge about ANNs and the problem to be solved. As ANN's complexity increases, designing ANNs manually becomes more diicult and unmanageable. Simulated evolution ooers a promising approach to tackle this problem. This paper describes an evolutionary approach to design ANNs. We call these ANNs designed by the evolutionary process as evolutionary ANNs (EANNs). In other words, EANNs refer to a special class of ANNs in which evolution is another fundamental form of adaptation in addition to learning (refers to weight training here). We use an evolutionary algorithm similar to evolutionary programming (EP) to evolve both archi-tectures and connection weights (including biases) of ANNs. Five mutation operators have been proposed for our evolutionary algorithm. In order to improve the generalisation ability of evolved ANNs, these ve operators are applied sequentially and selectively. We have also used validation sets in our studies to further improve generalisation. The evolutionary algorithm allows ANNs to grow as well as shrink during the evolutionary process. The evolutionary algorithm also incorporates the weight learning process as part of its mutation process. In a sense, our EANN system is a hybrid evolution and learning system. Extensive experimental studies have been carried out to test our EANN system. This paper will give some of the experimental results which show the eeectiveness of the system.