Effective exploration for action space in Profit Sharing

Abstract

Reinforcement Learning(RL) is one of the popular method in machine learning,which can adapt environment automatically. Most of RL such as Q-Learning guarantees to acquire optimum policy. The typical applications of this method are deciding the behavior of complex systems such as multi joint robots and multi-agent systems. However,it is difficult for Q-Learning to treat Partially Observable Markov Decision Process which exist in a lot of real environments. Profit Sharing(PS),which is one of the RL, can treats this process though the obtaining the optimum policy is not guaranteed. There is some proofs of acquiring rational policy in not only single agent systems but also multi agent systems. However,the PS needs a lot of trials to obtain the rational policy in the following environments: the agent requires many steps to get rewards, there are many actions which agent can execute. These problems occur in designing multi joint robot and a lot of agent in multi agent systems. In this paper , we proposed a new PS method which can learn fast in these environments and provide the availability of this method through numerical experimentations. The results show the proposed method can acquire the rational policy fast in comparison with the conventional methods.