Research On Improving BP Algorithm And Its Application

The artificial neural network is a hot research area in recent years, having many applications, including:information processing, materials science transportation, economy and still to expand. Also existing so many neural networks, but BP (Back-Propagation,referred to as BP) neural network still is the most widely used neural network. BP algorithm is based on error back propagation learning algorithm which has been widely used in variety of neural networks. This type of learning algorithm has a good generalization of nonlinear mapping capability and fault tolerance capability. However, due to the BP algorithm uses the steepest descent algorithm as learing-rule, there are some shortcomings such as:convergence is slow, easily to fall into local minima.At the same time, it has no unified theory as a guide to design its neural network structure. These defects have greately influenced the universality and application of BP algorithm.This paper analyzes the principle of BP algorithm, related improvements of BP algorithm. Based on these researchs,conceiving a compound error function which is based on the error rate for imprvoing BP algorithm’s defect of easily falling into local minima and a hierarchical dynamic adjustment mechanism of learning rate which is used to improve the convergence speed of BP algorithm. Finally, from the perspective of combining evolutionary algorithm with BP algorithm, based on improving genetic algorithm, constructed a BP neural network model based on improved genetic algorithm. The main work of the paper includes the followings: (1) This paper recalled the theoretical basis of BP neural network based on biological neural networks, overview of the theory of artificial neural networks. Emphasizing on the research of the BP neural network model, by detailly analysising of the process of derivation, made the limitations of the algorithm clear and lay the foundation for the next step to improve it.(2) As for the BP algorithm is easily to fall into local minimum, design a composite error function which is based on the error rate. This composite error function will implicitly take the particularity of the hidden layer into account, defining an error function of hidden layer and error rate which is used to measure the error. This error function will take the error rate as weight, adjust the weights of out layer and hidden layer according to the error amount. Finally, the feasibility of this algorithm is verified by experiments; The phenomenon that convergence rate is too slow exists for the BP algorithm. In this paper, propose a hierarchical dynamic adjustment of learning rate algorithm. This algorithm will config different learning rate for the hidden layer and output layer. These learning rates will be adjusted according to the error size and the change trendency of error which will effectively avoid the problem which is caused by the static learning rate in standard BP algorithm.(3) Combining the genetic algorithm which is good at global search with BP algorithm which has much strong local optimizing ability and constructing a BP neural network model based on improved genetic algorithm. This improved BP neural network model will optimize the BP neural network structure and the selection of weight by the usage of hierarchical coding, error-based fitness function, adaptive crossover operator.(4) Finally, applies the improved BP neural network model in the prediction of the coal gas content to Verify the validify of thie algorithm by comparative analysis with the standard BP algorithm.