EVOLVING BOTH THE TOPOLOGY AND WEIGHTS OF NEURAL NETWORKS∗
Zhengjun Pan, LISHAN NIE · International Journal of Parallel Emergent and Distributed Systems · 1996
Evolutionary algorithms (EAs) have been applied to designing and training artificial neural networks (ANNs), independently or in various combinations with other algorithms like back propagation and simulated annealing, which can be distinguished into the evolution of connection weights, of architectures and of learning rules. In this paper, we present an evolutionary approach to designing neural networks both for their topology and connection weights. For we directly represent a candidate network as a graph with weighted edges and vertices, the genetic operators can be designed to be quite natural and effective. For the mutation operator, we introduced the temperature of an individual on which the mutation based can protect fitter individuals while enlarge the search neighborhood of unfit ones. Although we do not use an extra procedure to train the connection weights of the networks in each generation, the experiments demonstrate that the algorithm can design an optimal network for some problems successfully and efficiently.