Reward-weighted DHER Mechanism For Multi-goal Reinforcement Learning With Application To Robotic Manipulation Control

In multi-goal reinforcement learning, an agent learns to achieve multiple goals using a goal-oriented policy, obtaining rewards from positions that have been achieved. Dynamic hindsight experience replay method improves the learning efficiency of the algorithm by matching the trajectories of past failed episodes and creating successful experiences. But these experiences are sampled and replayed by a random strategy, without considering the importance of the episode samples for learning. Therefore, not only bias is introduced as the training process, but also suboptimal improvements in terms of sample efficiency are obtained. To address these issues, this paper introduces a reward-weighted mechanism based on the dynamic hindsight experience replay (RDHER). We extend dynamic hindsight experience replay with a trade-off to make rewards calculated for hindsight experience numerically greater than actual rewards. Specifically, the hindsight rewards are multiplied by a weighting factor to increase the Q-value of the hindsight state-action pair, which drives the update of the policy to select the maximum action for the given hindsight transitions. Our experiments show that the hindsight bias can be reduced in training using the proposed method. Further, we demonstrate RDHER is effective in challenging robot manipulation tasks, and outperforms several other multi-goal baseline methods in terms of success rate.