Duplicated Replay Buffer for Asynchronous Deep Deterministic Policy Gradient

Off-Policy Deep Reinforcement Learning (DRL) algorithms such as Deep Deterministic Policy Gradient (DDPG) has been used to teach intelligent agents to solve complicated problems in continuous space-action environments. Several methods have been successfully applied to increase the training performance and achieve better speed and stability for these algorithms. Such as experience replay to selecting a batch of transactions of the replay memory buffer. However, working with environments with sparse reward function is a challenge for these algorithms and causes them to reduce these algorithms' performance. This research intends to make the transaction selection process more efficient by increasing the likelihood of selecting important transactions from the replay memory buffer. Our proposed method works better with a sparse reward function or, in particular, with environments that have termination conditions. We are using a secondary replay memory buffer that stores more critical transactions. In the training process, transactions are select in both the first replay buffer and the secondary replay buffer. We also use a parallel environment to asynchronously execute and fill the primary replay buffer and the secondary replay buffer. This method will help us to get better performance and stability. Finally, we evaluate our proposed approach to the Crawler model, one of the Unity ML-Agent tasks with sparse reward function, against DDPG and AE-DDPG.