The BP Neural Network Optimization Design Based On Improved GEP

Evolutionary algorithm and artificial neural network are two kinds of intelligent algorithm based on natural biological theory. They have been widely studied and applied because of their superior performance in solving certain problems. Evolutionary algorithms has become a powerful tool due to its good global search ability and Versatility of the algorithm for machine learning, optimization, function fields; and Artificial neural network is adaptive, nonlinear, parallelism, robustness and has the learning and associative functions. These features make it widely used in pattern recognition, signal processing, forecasting optimization and so on. However, with the rapid development and the research of two algorithms, they that are both based on the biological rules showed obvious trend of integration, and has formed a new research field -evolutionary neural network.Evolutionary algorithm (EA) is based on the theory of biological evolution proposed by Darwin. Through genetic operations of selection, crossover, mutation, the optimal individuals can be got by evolution. Gene expression programming (GEP) is a new evolutionary algorithm proposed by Portugal evolutionary biologist Ferreira in 2001. It can solve complex problems by simple coding to overcome the genetic algorithm (GA) and genetic programming (GP) problems. It also has the global search ability of evolutionary algorithms, but also very difficult to achieve the local optimal. BP neural network (BP-NN) is essentially a kind of gradient descent method, so it has strong local search ability. At the same time, as an effective algorithm for training multilayer networks proposed earliest, it has been widely used. But the disadvantage is the large amount of calculation, slow convergence speed, sensitive to initial value, easy to fall into local optimal solution.For GEP and BP problems, this paper studies deeply the characteristics of the two algorithms. Through the complementary advantages, the combination of the two algorithms was discussed. Firstly, the coding scheme to design the BP neural network using GEP was discussed, and neural network was represented by the chromosomes with the right domain and threshold array. Then, according to the characteristics of GEP neural network, the genetic operations of the GEP standard algorithm were improved adaptively, and to put forward a dynamic and evolutionary GEP algorithm (IGEP-BP). According to the problem that the traditional GEP designing can make the network structure loss of order, BP neural network orderly optimization algorithm based on GEP with structure domain (GEPO-BP) is proposed. Finally, the validity of the improved algorithm was tested by the simulation experiment.