Research On Multilayer Perceptron Learning Algorithm

Multilayer Perceptron is a sort of multilayer feed-forward single direct propagation network model.Because of its good nonlinear mapping ability, it is one of the basic models in the research andapplication of neural network at present, which has been widely applied to pattern recognition, imageprocessing, function approximation, optimization computation, optional prediction, adaptation controland so on. Multilayer Perception trained with BP algorithm often has a low convergence speed as anatural drawback, because it is based on gradient descent method which is only local searching. Whenapplied to an object function with many local minimums, it is not possible for BP algorithm to avoidbeing trapped in local minimum and to have a low converges speed. In a word, the research on BPalgorithm has become very important for a long time.The purpose of this design task is to study the algorithms of Multilayer Perceptron, and a new BPalgorithm is presented. Both BPWE algorithm (back-propagation by weight extrapolation) and TBPalgorithm (a three-term back propagation algorithm) are based on weight value adjusted. Considered toadd the proportional factor of the TBP algorithm into BPWE algorithm, it made the latter can adjustweight value by three terms too. A new BP algorithm, named TWEBP (the three-term weightextrapolation back propagation algorithm), is presented based on the two algorithm proposed just now.This new TWEBP algorithm is tested on three examples and the convergence behavior of the TWEBPand BP algorithm are compared. The results show that the proposed algorithm generally out-performsthe conventional algorithm in terms of convergence speed and the ability to escape from local minima.