Research On The Genetic Algorithm Improvement And Its Application In Neural Networks Control

As the development of science and technology, the scale and complexity of engineering systems become greater and greater. Then the intelligent control technology is developing rapidly as traditional control technique can't meet the engineering demand. It is widely used in variaty fields. Artificial neural network has very strong nonlinear approximation ability and fault tolerance and provides good control tools for complex control object. Genetic algorithm is not restrained by continuously differentiable conditons, and it has self-organization, self-learning ability and a strong global search ability.In this paper, genetic algorithm is improved specific to some of its shortcomings, then the improved genetic algorithm is used to optimize structure and weights of the neural network, and neural network algorithm based on the genetic algorithm is applied to the study of inverted pendulum control system. The main contents are as follows:1. An improved method of diploid genetic algorithm ignoring dominance and recessive of allele is provided for the standard diploid genetic algorithm is easy to fall into early convergence and has poor local search ability at final stage. It reduces the complexity of the algorithm and improves search accuracy. The genetic operation of traditional genetic algorithm is improved by imitateing the process of diploid biological reproduction, such as crossover of homologous chromosomes and gametes reoperation.2. Distant hybridization strategy and new method of crossover rate and mutation rate changes adaptively are provided based on multigroup cooperation evolution strategy. So the variation of individual is restrained by both number of genetic generation and fitness of evolution, making the search go into a larger solution space in the early stage of evolution, reducing the probability of the optimum model being destroyed in the later stage of evolution, and enhancing the global search ability of population.3. The definitions of improved individual distance and population distance are given based on traditional Hamming distance in consideration of the phenomenon of false distance for the different ranges of decision variables in the calculation of the distance between individuals. The search space is gradually narrowed by using the search area adaptive changing strategy on the basis of improved distance, and space consumption is reduced and search efficiency is improved. At the same time, the shrink of search space indirectly improves the utilization of protogene and enhances the local search ability of population. 4. To prevent the loss of good gene section after crossover, a linear crossover program related to individual fitness is designed on the ground of random non-linear crossover. By adjusting crossover rate and mutation rate adaptively, the premature convergence of the algorithm is effectively suppressed. And the improved genetic algorithm is used to optimize the parameters of nonlinear PID control and enhances control accuracy and robustness.5. For the structure of forward neural network set by experience, sensitive to the initial value and easy to get local optimum, using the improved genetic algorithms above to optimize the structure and weights of neural network is studied. And the idea is applicated in inverted pendulum control. Experimental results show that the method not only overcomes the randomness of selecting the network structure and the initial parameters, but also better guides the optimization directon and improves the generalization ability of neural networks.