Utilizing Hidden Observations to Enhance the Performance of the Trained Agent

Abstract

The frame-skipping strategy has been widely employed in deep reinforcement learning (DRL) technology to train an agent. Specifically, this strategy repeats the action determined by the agent for a fixed number of frames. It increases computational efficiency by reducing the number of inferences by making the action decision sparse. However, previously, these consecutive changes in frames during the frame-skipping were hidden and ignored from the environment and did not affect the agent's action decision. As a result, it can adversely affect the performance of trained agents, where the performance is more critical than computational efficiency. To alleviate these issues, we propose a new framework that utilizes these hidden frames during the frame-skipping, called hidden observation, to enhance the performance of the trained agent. The proposed framework retrieves all hidden observations during frame skipping. It then combines batch inference and an exponentially weighted sum to calculate and merge the outputs from hidden observations. Through experiments, we validated the effectiveness of the proposed method in terms of both performance and stability with only a marginal increase in computation.