Research On Improving Generalization Of BP Neural Network

With the deeply research on BP neural network, generalization is aroused widely debating of people as an important evaluation index. Generalization is the ability that neural network can identify the sets except training set, and it decides whether the neural network can be used or not in the production. The generalization of BP neural network is very sensitive, so it affects neural network's stability. In other words, it is easy to get into a local value and slow training speed, so the development of the BP neural network is restricted. People use simulated annealing, genetic algorithm and so on to improve topological structure and learning method. It can fit and forecast the neural network fast and accurately. By these ways the BP neural network's generalization can be improved.The factors about the generalization of the BP neural network are studied in this paper, for example sample space, initial weights, the parameters of network structure, and so on. Because BP neural network training speed is slow, network structure is hard to determine, and easy to get into a local value and etc, this paper proposes three kinds of different improved algorithms: GA-BP neural network improved algorithm based on PCA, determining automatically structure parameters of GA-BP neural network improved algorithm, determining automatically structure parameters of PCA-GA-BP neural network improved algorithm. GA-BP neural network improved algorithm based on PCA uses principle components analysis to reduce dimension, and uses genetic algorithms to optimize initial weights. It decreases network training time, and makes neural network to avoid getting into a local value. In order to confirm the number of BP neural network's hidden layers and the number of hidden layer neurons, the determining automatically structure parameters'GA-BP neural network improved algorithm is put forward by combining with genetic algorithm. Three layers BP neural network is changed by three or four layers BP neural network in this algorithm, and its prediction accuracy is increased. PCA-GA-BP neural network of auto-determining structure parameters based on GA-BP neural network of auto-determining structure parameters. It improves the sample space, initial weights and network structure parameters. So the speed of the BP neural network is increased. Through experiment data, three algorithms'generalizations are verified. Faced different sample data, the forecast accuracy of each algorithm is compared. This paper summarizes these algorithms'characteristic and adaptive conditions.