A modified learning algorithm for improving the fault tolerance of BP networks

The conventional back-propagation (BP) algorithm is not suitable for building fault tolerant neural networks, since it usually develops nonuniform weights. In this paper, a learning method to improve the fault tolerance in classification is therefore presented and a metric is devised to evaluate the performance. The new method is based on the BP algorithm. During the training, the magnitude of each weight is restrained from over-increasing. This modification enforces that the information be distributed across weights more evenly. Simulation results demonstrate that the modified algorithm leads to significant enhancement in the network´s ability to cope with internal hardware failures