Exploitation/exploration learning for MDP environment

Reinforcement learning is an effective learning in unknown environment, where a supervisor cannot support the learner. The agent needs a large number of trial-and-error interactions to find optimal behaviors. This leads to a serious problem if the agent is in a dynamic environment, because the agent cannot adapt to the new changed environment quickly. To overcome the drawback, we propose a new reinforcement learning method for quick adaptation. In the new method, the agent maintains both an exploitation (EI) strategy and an exploration (ER) strategy alternately for each state. In the El strategy, the agent tries to select the best action using the past memory of the agent. While in the ER strategy, the agent tries to identify the environment using an estimation of error and search new states from an unknown region in the state space. Using these two strategies, the agent can reduce the redundant searching in the state space. Experimental results show the agent yields a quick adaptation to unknown environments