A multi-agent approach to process the distributed computing of neural network

The modeling of many agents to collaborate training multi-layered perceptrons is designed to improve the learning speed and convergence property of NN (Neural Network), on the basis of the thought of multi-agent system. The algorithm system is composed of three kinds of heterogeneous agents: NN Agent, Server Agent, and GA (Genetic Algorithm) Agent. NN Agent and GA Agent respectively correspond to the method libraries to select data, algorithms, and communication mechanisms, and apperceive their internal and external states. Server Agent harmonizes NN Agent with GA Agent on the distribution of time and resource, and the management of the implementing agent queue, grounded on the knowledge-reasoning-based support library. The results of simulation studies and its application to inventory control are given, which show that this Algorithm are feasible and effective learning algorithms, which greatly improve the convergence property and avoid large computation complexity to some extent.