Comparative fault tolerance of generalized radial basis function and multilayer perceptron networks

A method for measuring fault tolerance is developed which provides the means to quantify the effect of large numbers of network faults without explosive computational complexity. The fault tolerance of two types of neural networks used for analog function approximation, i.e., the multilayer perceptron (MLP) and the generalized radial basis function (GRBF) network, is assessed. When standard gradient descent learning is used, the GRBF is considerably more fault tolerant than an MLP of the same size. When a fault tolerance enhancing training method is used, the fault tolerance of the GRBF improves substantially, while the fault tolerance of the MLP improves only marginally