Design of structural modular neural networks with genetic algorithm

Abstract

The back-propagation (BP) neural network and the radial basis function (RBF) neural network have been widely used in many engineering applications. In general, the BP neural network can better construct the global approximations to the input–output mapping, whereas an RBF neural network employs the exponentially decaying localized input–output mapping which can effectively model the large variation locally. In this paper, a structural modular neural network, by combining the BP neurons and the RBF neurons at the hidden layer, is proposed to construct a better input–output mapping both locally and globally. The use of genetic algorithm in searching the best hidden neurons makes the structural modular neural network less likely to be trapped in local minima than the traditional gradient-based search algorithms. An analytical function—the peak function is used first to assess the accuracy of the proposed modular approach. The verified approach is then applied to the strength model of concrete under triaxial stresses. The preliminary results show that the modular approach is more accurate than using the RBF neural network or the BP neural network alone.