Research And Application On BP Networks Based On Improved PSO

The traditional BP (Back Propagation) algorithm in BP network is the same as the rule of Widow-Hoff study, which belongs to gradient descent learning algorithms. The value of the amendments carries right along of the opposite direction of the gradient of the error performance function. Due to the complexity of high-dimension in practical application, the fact that BP neural network adopts traditional BP algorithm slows the training process and even leads to collapse of the network system. Since the traditional BP algorithm exists some shortcomings, there appears some BP neural network training learning algorithms based on genetic algorithm, particle swarm algorithm etc. particularly, the application of the Particle Swarm algorithms further accelerates the learning speed of the BP neural Network.PSO algorithm does not require a continuity of the objective function. In addition, its search can be characterized by parallel as well as an overall situation, thus its convergence features as simple and fast, and it no longer needs complex operations of selection, cross and mutation in GA (Genetic Algorithm) algorithm. However, for the complex issues of high-dimension, PSO algorithm tends to be easy premature and convergent, so it can not guarantee the convergence to the most advantage. Based on predecessors'research, this article not only conducts the research into the PSO algorithm convergence but also provides an modified algorithm called AMPSO (Particle Swarm Optimization with Adaptive Mutation), according to the fact that genetic variation can increase the diversity of the population. Led by the general truths in nature and the basic principles of physics community, in accordance with the changing environment, the AMPSO overcomes the problem of premature converg- ence phenomenon in the PSO algorithm.Finally, this article makes the AMPSO algorithm learning and training function of BP neural network. Besides, it builds BP neural networks on a basis of the AMPSO algorithm. What's more, the BP neural networks is applied to the technology of invasion detection. It turns out that based on the AMPSO algorithm, the iteration of BP neural Network times are less than other algorithms, and it has a higher average accuracy rate while it effectively solves the problem of slow convergence existed in traditional BP algorithm compared with the KDD99 CUP data based on different algorithms in BP network simulation experiments.