Modelling uncertainty in reinforcement learning