Multi-Agent Reinforcement Learning Theory And Its Application In The Price Reporting System Of Electric Power Industry

The study about agent and multi-agent is an area full of vitality. With the success of an agent's research, there has been a growing interest in the area of Multiagent System. Real systems are often complicated, huge and distributed while knowledge and resource of one agent is not enough, or inefficient, so study of multi-agent is developing rapidly, and the research has more challenge.The main work is about coordination and cooperation of agents in this thesis. How to make agent learn the skill of interaction with other agents by itself while giving attention to itself or the whole benefit is the problem we want to solve, which can be achieved through learning every games by combining reinforcement learning with game theory in the framework of stochastic game. A novel learning policy is proposed based Q-learning of single agent and other multi-agent reinforcement learning algorithms, which has been emulated in gridworld and trace experiments, and the results demonstrate its efficient and commonality. This novel learning policy also makes the research of multi-agent reinforcement learning extend Markov decision process(MDP) like environments to non-MDP-like environments. The application of quantum search algorithm in the search policy of the multi-agent's state and action, making use of quantum iteration theory sufficiently, seeks a new path different from traditional action selection method, which increases the parallelism of multi-agent reinforcement learning. The experiment results show that the new algorithm has a good convergence, and it makes a prior work for the future research in this field.This paper uses agent technique to resolve the problem of bid games in our power market when power companies are in different market environment, and presents a method of power bid games to determine the optimal bid.