Transfer in reinforcement learning