Reinforcement learning, or agent learning, has been a successful method in AI for learning policies, but lacks the capability to generalize to new distributions. Model-based RL, which uses generative models like GANs, could help agents to understand the world and generalize faster.