On the Optimum Method of Feedforward Multi-Layer Neural Network

Defining network average error as optimum objective function, weights and thresholds as design variable, which are rationally sorted, a new kind of real conjugate terraced optimum algorithm is studied. Compared with BP algorithm, the compute time is reduced and the precision is improved. A computing program about weights and threshold, based on high precision conjugate gradient optimum algorithm of multi-layer neural network, is put forward and programmed. The selecting method of rational construct is also pointed out. Through an application instance, its advantage and applying prospect is validated.