Consistent Meta-Reinforcement Learning via Model Identification and Experience Relabeling

Reinforcement learning algorithms can acquire policies for complex tasks automatically, however the number of samples required to learn a diverse set of skills can be prohibitively large. While meta-reinforcement learning has enabled agents to leverage prior experience to adapt quickly to new tasks, the performance of these methods depends crucially on how close the new task is to the previously experienced tasks. Current approaches are either not able to extrapolate well, or can do so at the expense of requiring extremely large amounts of data due to on-policy training. In this work, we present model identification and experience relabeling (MIER), a meta-reinforcement learning algorithm that is both efficient and extrapolates well when faced with out-of-distribution tasks at test time based on a simple insight: we recognize that dynamics models can be adapted efficiently and consistently with off-policy data, even if policies and value functions cannot. These dynamics models can then be used to continue training policies for out-of-distribution tasks without using meta-reinforcement learning at all, by generating synthetic experience for the new task.