Hindsight Trust Region Policy Optimization

Reinforcement Learning (RL) with sparse rewards is a major challenge. We propose Hindsight Trust Region Policy Optimization (HTRPO), a new RL algorithm that extends the highly successful TRPO algorithm with hindsight to tackle the challenge of sparse rewards. Hindsight refers to the algorithm's ability to learn from information across goals, including past goals not intended for the current task. We derive the hindsight form of TRPO, together with QKL, a quadratic approximation to the KL divergence constraint on the trust region. QKL reduces variance in KL divergence estimation and improves stability in policy updates. We show that HTRPO has similar convergence property as TRPO. We also present Hindsight Goal Filtering (HGF), which further improves the learning performance for suitable tasks. HTRPO has been evaluated on various sparse-reward tasks, including Atari games and simulated robot control. Results show that HTRPO consistently outperforms TRPO, as well as HPG, a state-of-the-art policy gradient algorithm for RL with sparse rewards.