Will it Blend? Composing Value Functions in Reinforcement Learning