Based On PSO-BP Network Learning Method Research

Along with the development of science and technology in the world, optimization’s technology has more and more important position. Optimization’s technology need to achieve the best effect with a short period of time. In recent years, the methods of network learning are attracting more and more attentions by a large number of scholars and research in many of the dynamic optimization algorithm.Back propagation is one of artificial neural network, which is a multi-layer feed-forward back propagation network. BP network learning compared with other algorithms, provides better ability of parallel processing and storing data, but also has been mentioned with low convergence and a long time of training, as well as the defects such as lower learning ability and generalization ability. In this paper, according to the theory of BP network learning problems, this paper proposes a momentum and adaptive learning rate adjustment of BP algorithm. The learning rate in the process of updating is updated through feedback adjustment, which can reduce training time. And the structure of BP network learning network is analyzed.BP network is easy to fall into local minimum point, in order to solve this question, particle swarm optimization algorithm is proposed. Particle swarm optimization algorithm is one of the optimization algorithms, which has good global searching capability. PSO have a good balance between convergence and diversity. In order to improve the algorithm of dynamic optimization and the performance of PSO algorithm, linear inertia weight is proposed. Using combination of PSO algorithm and BP algorithm, and which produce new PSO-BP network learning method. PSO-BP network learning method have higher stability and accuracy, the network learning ability and generalization ability, and promote a higher on the convergence with the characters of back propagation of BP network.In this paper, the main innovations can be summarized as follows:(1)Put forward the momentum and adaptive learning rate adjustment of BP algorithm. In order to improve the BP network learning fault-tolerant ability and ability in the process of optimization, put forward the momentum BP network of adaptive learning rate adjustment. The experimental results show that it improves the fault-tolerant performance, also improve prediction precision of BP network.(2)Give out linear inertia weight in PSO and the momentum and adaptive learning rate adjustment of BP algorithm in combination, which produce a new PSO-BP network learning method. Because of slow convergence speed of BP network, easily into the characteristics of the minimum, and the particle swarm algorithm parameters in the process of optimization problems, put forward the combination of both. In order to improve the global search ability of PSO algorithm, this paper proposes the methods of linear inertia weight, new PSO-BP network learning method is effective to improve the defects in the BP network learning method. Simulation experiments show that the convergence, accuracy and fault tolerance is better with PSO-BP, which has improved learning ability and generalization ability.(3)Improved the new PSO-BP network learning method, using back propagation of BP network learning mechanism. The characteristics of BP network have being better used. Experimental results show that improved PSO-BP before above had better stability, and have a really good learning ability and generalization ability, and have greater increasing in terms of convergence accuracy.In this paper, the main work can be summarized as follows:(1)Optimize BP network structure and PSO configuration parameter. In order to reduce the complexity of the network, and shorten the training time, this paper put forward only with a single hidden layer BP network. Due to the numbers of nodes influence the robustness of the neural network in the hidden layer of BP network, in order to achieve better prediction effect, this paper use Hecht-Nelson method to determine the number of nodes in the hidden layer.(2)There are five classical test functions, which make fitting the simulation experiments with the improved network learning algorithm. In the end, we analyze the data and comparison.