Reinforcement learning with rare significant events: direct policy search vs. gradient policy search

This paper shows that the CMAES direct policy search method fares significantly better than PPO gradient policy search for a reinforcement learning task where significant events are rare.