Research On Optimization Strategy Of BP Neural Network Training With MapReduce

BP neural network is one of the most popular artificial neural networks,which has been widely used in classification and approximation problem.BP neural network training is generally on a single machine and requires a long time as the increase the scale of the training data.Therefore,it is necessary to the parallelize the training of BP neural networks.MapReduce is a parallel computing model based on data-parallelism,which can be used for the parallelization of BP neural network training.However,the training results of a Map task can only achieve local convergence on the received input data slices,and the global weight matrix generated based on the average of weight matrix generaged by all Map tasks may not drive the training process to develop to the direction of global convergence.Therefore,solving these problems is the key to improve the training efficiency of BP neural network with MapReduce.To solve the above problems,two different approaches are proposed.The first method is based on the disordering of the training samples for each Map tasks by systematic sampling to generate new input for each them.It can speed up the Map tasks to generate global convergence weight matrices by using new input splits in the iterative training on the original weight matrix.Moreover,the weight matrix with the minimum training error produced by a Map task will be choosed as the initial weight matrix for the next round training.In the second approaches the Genetic Algorithm is applied in the Reduce task in the second method and the local convergence weight matrix of each Map task generated will act as the initial population of genetic algorithm.Therefore,the training process will develop to the direction of global convergence.The highest fitness weight matrix in current population will be choosed as the initial weight matrix of the next round of training of all Map tasks,which can speed up the training process furtherly.Simulation results show that the two proposed solutions can reduce either the training time of each round or the total number of iterations.At the same time,the two solutions have good extendibility as the growth of the cluster.