| Artificial Neural Network(ANN, for short) is a kind of information processing network, it ’s proposing is base on biological nervous system. Artificial Neural Network which is made up of many artificial neurons interconnected orderly and complicated with each other is a simplification and simulation of biological nervous system. Although the function of single artificial neuron is simple, the information transmission and processing ability of the huge Artificial Neural Network which consists of many artificial neurons interconnected complicated with each other.BP(Back Propagation) Neural Network is a typical Feedforward Neural Network. Owing to its simple structure and algorithms easy to achieve, BP Neural Network is one of the most studied and widely used Neural Network model. Benefit from many of its merits that include powerful nonlinear function mapping ability, parallel distributed processing ability, self-learning and adaptive ability, generalization ability, fault-tolerant capability, BP Neural Network is widely used in pattern recognition, information storage, image processing, fault detection, intelligent control, market analysis, business administration. However, its applica tion is limited because of its slow speed of convergence and its easy occurrence of fluctuation and shock. In addition, its performance is affected greatly by the initial training point and initial learning step length. So, the purpose of this paper is to increase the speed of convergence through related analysis and research, and to broaden its application and provide a guidance for the algorithm choices through the research in the influence of initial training point and initial learning step length to di fferent algorithms.The reasons for these weaknesses of BP Neural Network have been analyzed and studied through consulting lot’s of related document literature. On the basis, the theory and realization of overall variable step size algorithm and individua l variable step size algorithm have been researched and contrasted. Through simulation experiments of four examples and mathematical statistic analysis to the large amount of data, the final experimental conclusion has been made. Because the performance of BP Neural Network is influenced by its BP algorithm, transfer function, normalized interval, initial training point, initial learning step length and so on, the simulation experiments have been conducted under the condition of single variable and rep eated many times. Finally, in order to do scientific data analysis, the mean value and relative variance have been figured up.The main conclusions of this paper are as follows:(1) Both overall variable step length algorithm and individual variable step length can not only increase the speed of convergence greatly but also eliminate the fluctuation and shock when training the BP Neural Network.(2) overall variable step length algorithm is a proper steepest descent method, however, individual variable step length algorithm is not a proper steepest descent method.(3) In terms of speed of convergence, the ranking from fast to slow is individual variable step length algorithm of elastic search, overall variable step length with step length inheriting, ind ividual variable step length algorithm of coordinates rotation search and overall variable step length algorithm without step length inheriting.(4) In terms of the degree of influence aroused by initial training point, the ranking from small to big is overall variable step length algorithm without step length inheriting, individual variable step length algorithm of coordinates rotation search, overall variable step length with step length inheriting and individual variable step length algorithm of elasti c search.(5) In terms of the degree of influence aroused by initial step length, the ranking from small to big is overall variable step length with step length inheriting, individual variable step length algorithm of elastic search and overall variable step length algorithm without step length inheriting. |
PDF Full Text Request