The two facets of the exploration-exploitation dilemma

This paper proposes an algorithm to better solve the exploration-exploitation dilemma faced by model-less reinforcement learning agents. The main contribution is twofold: (1) The two facets of the exploration-exploitation dilemma are distinguished: in some cases, the agent faces a non-stationary environment, therefore it needs to choose the best moment to explore in order to adapt to the changes; in some other cases, the agent faces a relatively large state-action space, and it therefore needs to choose the most promising subset of states/actions to explore. In this two-facet framework, we compared the relative advantage and limitations of two previously proposed algorithms in difference situations. (2) We unified these two algorithms to produce the new algorithm which works fairly well in all testing situations.