PlaNet

Reinforcement Learning (5): Model-Based RL and World Models

From Dyna and MBPO to World Models, Dreamer, and MuZero -- how learning a model lets agents plan in imagination and reach expert performance with 10-100x fewer real interactions.