Hindsight policy gradients

Goal-conditional policies allow reinforcement learning agents to pursue specific goals during different episodes. In addition to their potential to generalize desired behavior to unseen goals, such policies may also help in defining options for arbitrary subgoals, enabling higher-level planning. While trying to achieve a specific goal, an agent may also be able to exploit information about the degree to which it has achieved alternative goals. Reinforcement learning agents have only recently been endowed with such capacity for hindsight, which is highly valuable in environments with sparse rewards. In this paper, we show how hindsight can be introduced to likelihood-ratio policy gradient methods, generalizing this capacity to an entire class of highly successful algorithms. Our preliminary experiments suggest that hindsight may increase the sample efficiency of policy gradient methods.