Abstract:

Reinforcement learning is  not applicable concerning large state-actions, since that its convergence speed increases exponentially with the number of dimensions of state-action space. In many situations, this problem partially can be solved  by utilizing a cooperation relationship among agents. An evolutional algorithm was put forward, which could rapidly find the effective updating of state-action pairs by the evolutionary operators such as reproduction as well as die out. Simulations proved that the algorithm performs was better than present multiagent cooperation learning algorithms.

Key words:  multiagent system, cooperation learning, evolutionary algorithm