model-based reinforcement learning (MBRL) holds promise for data-efficiency
by planning with model-generated experience in addition to learning with
experience from the environment. However, in complex or changing environments,
models in MBRL will inevitably be imperfect, and their det