In reinforcement learning, planning is the process whereby an agent imagines future trajectories or plans using a learned or given model. This model-based reinforcement learning (which combines planning with learning) outperforms several other approaches, including model-free methods, zero and few-shot learning, and strategic thinking. However, there are a wide variety of ways to implement planning. In this paper, Hamrick et al. help practitioners understand what form of model-based reinforcement learning may work best for a specific problem by studying the performance of MuZero. They find that although planning does not induce strong generalization, it is useful for policy updates and for providing a more useful data distribution