Efficient Off-Policy Meta-Reinforcement Learning via Probabilistic\n Context Variables

Abstract

Deep reinforcement learning algorithms require large amounts of experience to\nlearn an individual task. While in principle meta-reinforcement learning\n(meta-RL) algorithms enable agents to learn new skills from small amounts of\nexperience, several major challenges preclude their practicality. Current\nmethods rely heavily on on-policy experience, limiting their sample efficiency.\nThe also lack mechanisms to reason about task uncertainty when adapting to new\ntasks, limiting their effectiveness in sparse reward problems. In this paper,\nwe address these challenges by developing an off-policy meta-RL algorithm that\ndisentangles task inference and control. In our approach, we perform online\nprobabilistic filtering of latent task variables to infer how to solve a new\ntask from small amounts of experience. This probabilistic interpretation\nenables posterior sampling for structured and efficient exploration. We\ndemonstrate how to integrate these task variables with off-policy RL algorithms\nto achieve both meta-training and adaptation efficiency. Our method outperforms\nprior algorithms in sample efficiency by 20-100X as well as in asymptotic\nperformance on several meta-RL benchmarks.\n