SPLID: Self-Imitation Policy Learning through Iterative Distillation

Goal-Conditioned continuous control tasks remain challenging due to the sparse reward signals. To address this issue, many relabelling methods like Hindsight Experience Replay have been developed and bring significant improvement. Though relabelling methods provide an alternative to an expert demonstration, the majority of the relabelled data are not optimal. If we can improve the quality of the relabelled data, the sample efficiency, as well as the agent performance, should be improved. To this end, we propose a novel meta-algorithm Self-Imitation Policy Learning through Iterative Distillation (SPLID) which relies on the concept of $\delta$-distilled policy to iteratively level up the quality of the target data and agent mimics from the relabeled target data. Under certain assumptions, we show that SPLID has good theoretical properties of performance improvement and local convergence guarantee. Specifically, in the deterministic environment, we develop a practical implementation of SPLID, which imposes $\delta$-distilled policy by discriminating First Hit Time (FHT). Experiments show that SPLID outperforms previous Goal-Conditioned RL methods with a substantial margin.