On the use of predefined regions to minimize the training and complexity of multilayer neural networks

The flexibility of the multilayer perceptron (MLP) in conjunction with the back-propagation training algorithm has been instrumental in the resurgence of interest in the study of artificial neural networks. The MLP can map any distributed set of arbitrarily shaped regions, and therefore functions as a universal mapping network. However, it is demonstrated that the flexibility of the design caused a penalty in terms of training time. The MLP is compared with two other networks that contain predefined regions, the regions being either an ellipsoid or a sphere. The three networks had the same number of adjustable parameters to enable a fair comparison and each of the three networks was trained using two test patterns. In both tests the MLP performed poorly, calling into question its usefulness in a practical environment. In contrast the other networks performed well, acknowledging the presence of disjoint regions early in the training sequence, with the network based on ellipsoids performing marginally better than the one based on spheres. It is concluded that the radial basis function with ellipsoids gives a better performance than the MLP