A Generalized Circle Agent Based on the Deep Reinforcement Learning for the Game of Geometry Friends

摘要

In this paper, we study creating a generalized circle agent based on deep reinforcement learning for the game of Geometry Friends. We use the same setups proposed by another paper studying the game of Geometry Friends. The proposed deep reinforcement learning-based agent is trained with the Rainbow algorithm, which is a combination of solutions to different problems in the field of reinforcement learning. Our trained agent successfully completes all setups and shows a significantly higher performance over the agent trained in the previous study. In addition, performance of our agent is superior compared to human performance in the same setups. The agent demonstrated a performance pattern similar to that of human, i.e., the setups spent longer time are the same for both.