A new action selection policy in reinforcement learning problems based on fuzzy mappings

In this paper, a new policy in action selection processes related to reinforcement learning problems is presented. This policy in fact is a fuzzy mapping that attributes probabilities of selection to actions proportional to agent s view. Agent s view in this method inculcates to system by a control parameter named ξ that tunes and adjusts very simpler than τ in boltzman softmax method. Intuitiveness and interpretability of the parameter ξ because of using fuzzy system gives us opportunity to contribute the human knowledge in the action selection process. Better performance and more rapid convergence also are two other significant causes for superiority of proposed method.

Keywords:

Action Selection Policy , Softmax , Fuzzy mapping , n-Armed bandit problem