This paper presents simple techniques to improve the learning rate of the RL algorithm in a task with large number of states that the algorithm cannot fully explore by going through all possible states. The problem with the standard RL algorithm is that varying the ε parameter in the ε-greedy policy is not sufficient to improve its learning performance. The techniques allow the algorithm to focus on exploring the local path to eventually obtain more rewards and occasionally switching to another path to avoid being trapped in the local region of the search space. In order to investigate the effectiveness of the techniques, the minimal consistent subset identification (MCSI) problem is used as a test problem. The paper concludes by comparing the size of the identified subset obtained from the standard RL algorithm and the proposed algorithm along with those of the standard MCSI method.
