Structure Optimization Of Feed-forward Neural Network Based On Genetic Algorithm

BP networks, although there are simple advantages, but its essence is based on thegradient descent algorithm, BP neural network may be trapped into local minimum, and thushinder BP neural network learning, at the same time, network generalization should bereduced. In addition, when BP network to solve more complex problems, to design artificialnetwork structure is not easy to achieve the right results. Genetic algorithm is a branch ofartificial intelligence, widely used, has a strong global income and optimize computing powercable, does not depend on the gradient information, suitable for solving highly complexnonlinear problems, moreover, the problem can be massively parallel distributed processing,can solve the problem of BP network architecture design.Paper studies the structure of neural network to determine the basic methods, discussesthe conventional neural network structure to determine the problems and deficiencies, wereinvestigated additional momentum, adaptive learning rate method, conjugate gradient method,Levenberg-Marquardt (LM) algorithm in neural network design applications, zoom proposedgenetic algorithm and BP algorithm combines neural network structure design methods.Taking into account the presence of BP neural network training is slow, localoptimization problem, the genetic algorithm optimization of the overall search strategy andcalculation does not depend on gradient information, especially for highly complex nonlinearproblems; because genetic algorithm parallelism, makes it more suitable for massivelyparallel distributed processing, combines the advantages of genetic algorithm, given based ongenetic algorithm optimization neural network structure feasibility, And gives thecombination of the two specific methods, Simulation results show the effectiveness of thegiven method.Papers on genetic algorithms using hybrid coding, design an adaptive genetic approaches,let the crossover probability and mutation probability dynamically adjust the process ofevolution. then, for function approximation problem using the improved genetic algorithm tooptimize the structure of feed-forward neural networks, improve the network convergencespeed and network generalization ability, avoid local minima problems.