Research On Training BP Neural Network By Improved Particle Swarm Optimization Algorithm

BP neural network is widely used in many areas including pattern recognition, artificial intelligence, signal processing and automatic control. It is one of the most widely studied and applied artificial neural network. In many application areas, BP neural network exposed some problems such as slow convergence rate, sensitive dependence on initial conditions, the existence of local minimum point, the difficulties of determining the number of hidden layers and the number of neurons. In order to overcome the defects of BP neural network algorithm, this paper presents an improved method of simplified particle swarm optimization algorithm to train the weights and thresholds of BP neural network and determine the structure of BP neural network. The main work of this paper is as follows:1) A novel simplified particle swarm optimization algorithm (SIWSPSO) based on stochastic inertia weight is proposed to avoid premature convergence and low search precision of the standard PSO. The proposed algorithm is based on the Simplified PSO which removes the velocity, where inertia weight is represented as stochastic variable, learning factor uses asynchronous change strategy, and the individual extreme value of each particle is represented by the mean of the individual extreme values of all particles. Through simulation on several typical benchmark functions and F test, it shows that there is a significant improvement for SIWSPSO in searching speed, convergence accuracy and robustness, compared with the existing improved algorithms. Meanwhile, SIWSPSO effectively overcome the drawback of discretization, low convergence speed in the later period of the optimization, premature convergence of standard PSO.2) Combining simplify particle swarm optimization algorithm and BP neural network to build a new maxing algorithm, and use which to train the network’s weights(include threshold) and build the structure of network simultaneously. In this new mixing algorithm, dimensions of the particles is equals the sum of the number of neural network weights and threshold number plus the number of hidden layers in network. Through simulation results show that this mixing algorithm has high accuracy and precision fitting ability in cope the problem of classification in pattern recognition. This algorithm can not only avoid the BP neural network sensitive to initial values, slow convergence problem of the existence of local minima, but also can get a good network topology without using empirical formula to gradually analysis.