Abstract: It is difficult for humans to obtain new knowledge, but we can acquire new knowledge by imitating others’ behaviors. Inspired by such human characteristics, we propose a deep reinforcement learning method called \textit{memory reinforcement learning}. Our approach leverages the technique of experience replay and a replay buffer architecture. We manually create stable action transition sequences (prior knowledge) and store these transitions in the replay buffer at the beginning of training. These hand-crafted transition sequences enable us to avoid random action selections and a local optimal policy. Consequently, our method can acquire stable control efficiently. Experimental results on a lane-changing task of autonomous driving indicate that the proposed method can efficiently acquire stable control.
2 Replies
Loading