Abstract
In this paper, we propose a swarm reinforcement learning method for dilemma problems of multi-agent tasks in which it is difficult for agents to learn cooperative actions. In the proposed method, multiple sets of the agents and the environments, which are called learning worlds, are prepared and each agent in each world learns through exchanging information with agents in all other worlds. In order to acquire the cooperative actions, we propose a method of information exchange in which the agents in all learning worlds share the state-action values which are estimated to be superior for taking cooperative actions. The proposed method is applied to the N-persons iterated prisoner's dilemma and the Tragedy of the commons that are typical dilemma problems, and its performance is evaluated by investigating the learning processes and results.
