An enhanced parallel backpropagation learning algorithm for multilayer perceptrons

Backpropagation learning algorithm for multilayer perceptrons (MLPs) has disadvantages of slow convergence and easily being trapped into local optimum. Inspired by efficient global searching ability of particle swarm optimization (PSO), a novel PSO based backpropagation learning algorithm (PSO-BP) is proposed. At first, training procedure for MLPs is formulated as nonlinear optimization problem that can be processed by PSO. Then, combination of PSO learning and BP learning is used to train architecture and weights of MLPs. By using particle update equations, PSO learning provides optimal architecture and weights under the condition of input and desired output that are known. BP learning provides optimal weights of MLPs under the condition of given architecture and initial weights that have been set by PSO learning. The proposed learning algorithm is applied to load forecasting in electric power system. Test results show that the proposed algorithm can effectively avoid to be trapped into local optimum and has faster convergence speed than BP algorithm.