Function approximation using the cooperative PSO neural network

Among the methods of training weights of neural network, BP is the most popular for its reasonable basis. But there still are some unsolved problems, such as the training result being influenced greatly by the order of samples, local optimization and the slow learning speed etc. PSO (Particle Swarm Optimizer) is often used to replace the BP method to speed up the learning and reduce the probability of failing into local optima. According to the topologic structure of network, a pattern of cooperation called FLsplit is prompted, which will be convenient for parallel computation. And a particular individual named BParticle is involved in the swarm, to stabilize the learning result and decrease the computation cost. The performances of BP, standard PSO and FLsplit are analyzed through the experiments on two functions, and the efficiency of FLsplit is addressed.