Research On Cooperative Multi-agent Reinforcement Learning With Pursuing Problem

Among the methods of machine learning, the reinforcement learning is the most popular, which has succeeded in single agent system. Today's works are focused on the learning in multi-agent system, where the complexity and uncertainty make the learning more difficult. Game theory is the framework to investigate the interaction of several players. When combined with the Markov decision process, it provides a new formalization suitable for multi-agent system. That is stochastic game concerning the interactive learning system of multi-agent.There have been many works in this field, which are learning in adversarial system, non-cooperation system, full-cooperation system and repeated game respectively. The learning methods in full-cooperation system and rational-cooperation system are considered here.In the full-cooperation system, the learning method of single agent is adopted. The bias technology and information share are considered further to speed up the learning, and a policy-shared learning method based on the prior-knowledge is prompted.Besides the algorithm, the neural network realizing the learning is also important for its performance of speed and convergence. BP is one of the most popular training methods for multi-layer neural network. But there still are some unsolved problems, such as the training result being influenced by the order of samples, local optimization and the slow learning speed etc. A method using cooperative particle swarms optimization is proposed to replace the BP method, so as to optimize the weights of network quickly and globally.The works on rational-cooperation system are not so many. The critical techniques discussed here contain two aspects. One is the selection and computation of objective function to be learned, and the other is the coordination...