Zero-Shot Task Generalization With Multi-Task Deep Reinforcement Learning