Model-Based Stabilisation of Deep Reinforcement Learning