Maximizing fault tolerance in multilayer neural networks

Good fault tolerance is desired in many control and automated systems. It has been long claimed that neural networks are fault tolerant, i.e., they can continue to operate after sustaining partial damage. However, very little evidence shows that neural networks evolved from general learning algorithms really possess such merit. In this paper, the authors examine a learning method that intends to maximize the fault tolerance. The method is based on the well-known backpropagation learning algorithm. During the training, each neuron is given a small probability to have a simulated failure. This modification enforces that the computation be distributed to different computing elements in the network and thus maximizes the fault tolerance. Neurocontrollers developed using the new and conventional learning methods for pole balancing are compared