The Research And Application Of Neural Network Based On The Maximum Likelihood Method

At present, the theory and application studies of the neural network have got a very big development, and already permeated to nearly all project application fields. The BP neural network is one kind of multilayer feed forward neural network, named from back-propagation algorithm, which is the adjustment regulation of the networks' weight. According to the statistics, 80%-90% of the neural network models adopted BP neural networks or its modified versions. People have proved that in theory the various feed-forward neural networks can approximate the [ ]L2 R functions or more general nonlinear functions in any accuracy. However, how to gain the optimum network weights in real applications is not easy. As for BP algorithm, it usually fails to overcome the difficulties, such as slow convergence speed, weak generalization abilities and so on. Especially, when the dataset containing noises is used for training network weights, BP algorithm usually converges slowly and even fails to converge. Therefore, how to effectively enhance the anti-noise abilities and generalization performance of the BP network is a very important research topic.The learning rules of classical BP neural networks are gained from the energy function of least mean error criterion, and the training procedure of network weights is guided by the energy function. First, in this study, the weaknesses of classical BP neural networks are analyzed from the energy function. Then to overcome the weaknesses, the maximum likelihood method is introduced into the traditional BP neural network. Second, a novel energy function is proposed and with the maximum likelihood method the new learning rules are obtained to train the network weights. Many simulated experiments are conducted to test the performance of our new method. The experimental results demonstrate that our new method has the better robustness and generalization abilities. Furthermore, the new method is used to some real applications. The experimental results also show its advantages.