The Role of Uncertainty in Model-based Reinforcement Learning