Meta-Reinforcement Learning Based on Self-Supervised Task Representation Learning

Meta-reinforcement learning enables artificial agents to learn from related training tasks and adapt new efficiently with minimal interaction data. However, most existing research is still limited narrow task distributions that are parametric stationary, does not consider out-of-distribution during the evaluation, thus, restricting its application. In this paper, we propose MoSS, a context-based algorithm based on Self-Supervised representation address challenge. We extend meta-RL broad non-parametric which have never been explored before, also achieve state-of-the-art results in non-stationary tasks. Specifically, MoSS consists of inference module policy module. utilize Gaussian mixture model for imitate variations. Additionally, our online adaptation strategy agent react at first sight change, thus being applicable exhibits strong generalization robustness out-of-distributions benefits reliable robust representation. The built top an off-policy RL entire network trained completely ensure high sample efficiency. On MuJoCo Meta-World benchmarks, outperforms prior works terms asymptotic performance, efficiency (3-50x faster), efficiency, diverse distributions.