Abstract:

To improve the efficiency of BP neural network (BPNN) training, we propose an approach with MapReduce based on the local weight matrix evolution. The local weight matrices produced by Map tasks based on its input data splits are passed to Reduce tasks as the initial population of a genetic algorithm. After the evolution of the current population, the weight matrix with highest fitness is selected as the initial weight matrix in the next turn of training. The training does not stop until the weight matrix is convergent on the whole sample data. Experiments show that the proposed approach can maintain the global convergence of a BPNN on its whole training sample data, and improve the efficiency of BP-NN training with MapReduce.

Key words: BP neural network, genetic algorithm, MapReduce, convergence