Improvement To Artificial Neural Network And Its Application In Economics Analysis

Artificial neural network (ANN) model is a kind of intellectual algorithm that is set up based on biological neural network. ANN has the properties such as large scale parallel distributed processing, fault-tolerance, self-organized learning and self-adaptivity, non-linear approximation and classification etc. It is a new method to help to break through the existing bottleneck of science and technology and further investigate the non-linear and complicated problems. The neural network algorithm has already been applied successfully in many fields such as data classification, patternrecognition, economic forecasting, to name a few.In recent years studies on neural network algorithm are in the ascendant, especially in how to increase the speed of convergence and how to avoid minimum. Experts have put forward many effective methods. But there still is a difficult problem in classical algorithm: how to determine the rational neural networks architecture namely how to determine the number of hidden layers and the number of hidden nodes. In practical application, the network architecture is general determined with more subjectivity, experiences and less objectivity, rational standard. In this paper the number of input layer nodes is determined by principal component analysis method, the number of hidden layer nodes is optimized through merging and deleting and a new algorithm on the basis of result feedback (FBBP) is extended. The detailed is as following.In Chapter 1 we introduce the development and the basic algorithm of traditional BP network briefly. The problem that how to determine the network architecture is put forward. In Chapter 2 in order to avoid the too complicated network architecture caused by too many input nodes we reduce the input nodes with the principal component analysis method and factor analysis method. In Chapter 3 we firstly optimize the connection of network architecture though genetic algorithm, then weintroduce the concepts of relation degree and scatter degree. The nodes of hidden layer are merged and deleted according to the value of relation degree and scatter degree. In Chapter 4 we extend a two-layers network algorithm to usual K-layers network. In Chapter 5 a neural network model is set up with the optimized input layer and hidden layer. We forecast the consumption ratio of Shandong Province and evaluate the performance of some cement mills. The result shows that the improved algorithm is better than classical algorithm no matter in convergence speed or in predicting the precision.