A Novel Learning Algorithm of Back-Propagation Neural Network

Abstract

Standard neural network based on back-propagation learning algorithm has some faults, such as low learning rate, instability, and long learning time. In this paper, we introduce trust-field method and bring forward a new learning factor, meanwhile we adopt Quasic-Newton algorithm to replace gradient descent algorithm. Three algorithms are utilized in the novel back-propagation neural network. Thus the neural network avoids the local minimum problem, improves the stability and reduces the training time and test time of learning and testing. Two concrete examples show the feasibility and validity of the new neural network.