Optimize The Design Of BP Neural Network Based On Improved Artificial Fish Swarm Algorithm

Back-propagation (BP) neural network，as one of artificial neural networks, is currentlyused widely and grows fastest. According to statistics,90%of artificial neural networkmodels use BP neural network or its variations in the practical application. However, BPneural network is so easy to fall into local minima that it can’t find the global optimum.Therefore, two improvements are proposed in this paper. First, the paper uses artificial fishswarm algorithm to optimize the weights and the thresholds of BP neural network. Second,modular BP neural network is applied, which uses multiple neural networks instead of aneural network.Artificial fish swarm algorithm is an artificial intelligence optimization algorithm, whichis proposed by Li Xiaolei in2002. We often use artificial fish swarm algorithm to optimizethe weights and thresholds of BP neural network, because it can quickly find the globaloptimum and overcome the shortcomings of BP neural network. This paper makes threeimprovements for artificial fish swarm algorithm. First, the paper proposes fish initializationmethod based on cluster, which produces the fish that embody uniform distribution in theoptimization space so as to find the optimal value quickly. Second, three improvements forbasic behaviors of the fish are given. We introduce the information of the best individual inhistory and use heuristic algorithm for optimization. Finally, the escape behavior is brought infor fish, and in order to avoid the fish to fall into local minima, the paper uses crossover factorand variation factor of genetic algorithm.In the practical application, the BP neural network structure is very complicated whenthe problem has a lot of categories, which results in more misclassifications and lowerclassification accuracy if a suitable one is not established. So the paper proposes modularneural network with multiple neural networks instead of a neural network, and every data setis divided into two categories by a neural network, so the complexity of algorithm is reduced.Greatly multiple neural networks deal with dataset in the meantime and greatly reducestraining time of the network. Moreover, each neural network only deal with a binaryclassification problem, so the testing accuracy of the network are improved effectively.