Transfer in Deep Reinforcement Learning Using Successor Features and Generalised Policy Improvement