Generalization to New Actions in Reinforcement Learning