Re-exploration of ε-greedy in deep reinforcement learning

Abstract

This paper presents re-exploration as a method for improving the existing method for balancing the exploration/exploitation problem integral to reinforcement learning. The proposed method uses a ε-greedy method called “decreasing epsilon,” which reiterate the method after a certain period of episodes in the middle of the learning. The experiment was conducted using Turtlebot3 simulation under the Robot Operating System (ROS) environment. The evaluation involved comparing the existing method, which is pure exploitation (totally greedy), conventional ε-greedy method and proposed method, which is decreasing-epsilon with the re-exploration method. The preliminary results indicate that applying re-exploration method is easier to implement and yet able to improve the reward obtained with in shorter time (episode) compared to the conventional method.