Application Research Of BP Neural Network Structure Optimization Based On Genetic Algorithm

Genetic algorithm is a kind of intelligent search algorithm. It has some advantages, such as having global optimization performance, versatility and being suitable for parallel processing. Artificial neural network has a wide range of applications in terms of function approximation, pattern recognition, and data compression. While it is difficult to choose the number of hidden-layer nodes, initial weights and thresholds. The neural network research combining with genetic algorithm is a hotspot in recent years. The study that genetic algorithm optimizing the BP neural network structure, mainly uses the global search function of genetic algorithm.Firstly, Because of the distribution imbalance of the initial population that genetic algorithm generated randomly, this paper proposed an equidistant sampling method for initial population. In the study, the optimal individual is easily searched. What is more, this way accelerates the search speed of genetic algorithm.Secondly, Fitness function is treated as the only standard in judging the individual merits. Later stage in evolution, it is difficult to keep diversity of population. This paper adopted a new method for function scaling. It takes full use of best fitness, average fitness and evolution generations of the population to adjust fitness function. The purpose is to limit competition in early evolution, and later encourage competition. Thirdly, after a thorough theoretical study of genetic algorithm, this paper given improvement measures about standard genetic, such as using grouping method in roulette selection, adopting adaptive probability of crossover and mutation. This way can prevent some individuals from replacing quickly other individuals. In addition, because there has a smallest "difference" between the current and global optimal individual, this paper added translation operator of the optimal individual. It is sure the optimal individual can able to move effectively to approximate the optimal solution. And the improved genetic algorithm was tested using some classic genetic functions.Finally, in order to unity the standard of BP network initial weights and threshold, this paper presents BP network model, which has a maximum limit hidden-layer nodes. Eventually the optimized network is a sub network of initially BP network. Simulation experiment was made for the modified genetic algorithm. Compared with other methods, the improved genetic algorithm was proved to improve mapping capability of the BP network generalization. And it makes the BP network has stronger learning ability, higher convergence speed and efficiency. The structure of neural network optimized is more reasonable. Furthermore, it has higher precision in classification and approximation.