Sequential optimistic ad-hoc methods for nonstationary multi_armed bandit problem

Abstract

One of the common ways for showing the trade_off<br><br>between exploration_exploitation in reinforcement learning<br><br>problems is the multi_armed bandit problem. In this paper<br><br>we consider the MABP in a nonstationary environment which<br><br>features change during the period of learning. The<br><br>represented learning algorithms are intuition based solutions<br><br>to the exploration_explotation tarde_off that are called ad<br><br>hoc method. These methods include action_value methods<br><br>with e-greedy and softmax action selection rules, the<br><br>probability matching method and finally the adaptive pursuit<br><br>method. For producing near optimal results we change the<br><br>ad hoc methods to sequential optimistic ad hoc methods<br><br>which provide us completely better results.