An adaptive method of training multilayer perceptrons

A training method is proposed that adaptively select the sensitivity index of the risk-averting training criterion to suit the function under approximation and the training data used, when the measurement noises are unbiased. The proposed adaptive training method using a succession of risk-averting criteria is able to tune to the size of and include fine features and under-represented segments of the function. Numerical examples are given illustrating the efficacy of the proposed adaptive risk-averting training method. Most important perhaps, the new training method seems capable of avoiding poor local extrema of the selected training criterion.