Regularized Reinforcement Learning

人工智能的关键挑战之一是顺序决策问题， 代理人必须选择行动，以最大化选定的结果。强化学习提供了丰富的 数学和算法框架来解决这个问题，但部署强化 实践中的学习方法仍然落后于其他机器学习范例。在 科学水平，限制因素之一是缺乏对过度拟合问题的重视， 强化学习我们建议通过制定严格的治疗方案来解决这一问题， 正则化方法用于强化学习。开发的方法的有用性将在实际的顺序决策问题的背景下，在医疗保健和辅助机器人领域进行调查。 通过提供有效泛化的原则性工具，这些贡献预计将对我们在实际应用中部署强化学习解决方案的能力产生重大影响。

One of the key challenges of artificial intelligence is the sequential decision-making problem, whereby an agent must select actions, to maximize a selected outcome. Reinforcement learning provides a rich mathematical and algorithmic framework for tackling this problem, yet the deployment of reinforcement learning methods in practice continues to lag beyond that of other machine learning paradigms. At the scientific level, one of the limiting factors is the lack of attention paid to the problem of overfitting in reinforcement learning. We propose to tackle this problem by developing a rigorous treatment of regularization methods for reinforcement learning. The usefulness of the methods developed will be investigated in the context of practical sequential decision problems in the fields of health-care and assistive robotics. The contributions are expected to have a major impact on our ability to deploy reinforcement learning solutions in practical applications, by providing principled tools for effective generalization.