Generalizing Reinforcement Learning to Unseen Actions