CAREER: Non-asymptotic, Instance-optimal Closed-loop Learning

职业：非渐近、实例最优闭环学习

• 批准号: 2141511
• 负责人: Kevin Jamieson
• 金额: $ 50.69万
• 依托单位: University of Washington
• 依托单位国家: 美国
• 项目类别: Continuing Grant
• 财政年份: 2022
• 资助国家: 美国
• 起止时间: 2022-06-15 至 2027-05-31
• 项目状态: 未结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=2141511&HistoricalAwards=false
• 关键词: CAREER; Non; asymptotic; Instance; optimal

Machine Learning and Artificial Intelligence can recognize and exploit hidden patterns in data in order to predict future outcomes in applications ranging from content recommendation to personalized medicine. However, there are many problem areas where collecting the data is time-consuming (e.g., cells need to grow in lab environments) or expensive (e.g., special materials or expert opinions are required). Ideally, in order to reduce the amount of data needed to reach conclusions, already-collected data can be leveraged to guide the selection of future measurements in a closed-loop manner. While the behavior and benefits of some closed-loop data collection strategies are well understood in simple settings, this family of strategies is not commonly employed in real-world scientific laboratories or in medical trials due to a lack of predictability and accuracy of the outcomes. This project seeks to make foundational contributions to the understanding of closed-loop learning strategies with a view towards designing new data-collection strategies that are both effective and reliable. In practice, this may lead to requiring fewer patients in a clinical trial or to halving the time to identify a disease-curing drug. The investigator also plans to engage with high-school students and machine-learning enthusiasts alike to increase their level of awareness around data collection -- for instance, how even a simple survey, if not carefully designed, can result in privacy violations, demographic under-representation and bias of many forms, all of which may lead to inaccurate conclusions.For many problems of interest in closed-loop learning, prior art has focused only on minimax optimality, where the sample complexity of the worst-case problem instance is minimized. This approach leads to algorithms that are significantly inferior on "easy" or benign instances that may occur in nature but which are far from adversarial. In contrast this project will study the fundamental limits of instance-optimal sample complexity for problems of interactive learning and reinforcement learning in the Probably Approximately Correct (PAC) setting. The insights to be gained will be applied towards the design of algorithms that automatically adapt to the intrinsic difficulty of the particular problem instance being faced, be it benign or not. The proposed approach is motivated by the observation that the instance-optimal sample complexity decomposes into an asymptotic term, which is by now well characterized, and a moderate-confidence term, which is known to dominate the asymptotic term for all practical purposes. As the properties of the latter term are still poorly understood, lower bounds for it will be constructed together with algorithms that nearly achieve them. Such results will lead to algorithms that greatly reduce the overall instance-optimal sample complexity and vastly improve upon state-of-the-art algorithms that tend to cater to worst-case scenarios. The efforts will initially focus on structured linear bandits and reinforcement learning in the tabular and linear-function approximation settings. While these paradigms are of wide applicability to practice, they also have enough complexity to allow insights to be extrapolated to more generic closed-loop learning paradigms.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.

机器学习和人工智能可以识别和利用数据中的隐藏模式，以预测从内容推荐到个性化医疗等应用程序的未来结果。然而，有许多问题领域，收集数据耗时（例如，细胞需要在实验室环境中生长）或昂贵（例如，需要特殊材料或专家意见）。理想情况下，为了减少得出结论所需的数据量，可以利用已收集的数据以闭环方式指导未来测量的选择。虽然一些闭环数据收集策略的行为和好处在简单的环境中得到了很好的理解，但由于缺乏结果的可预测性和准确性，这一系列策略在现实世界的科学实验室或医学试验中并不常用。本项目旨在为理解闭环学习策略做出基础性贡献，以期设计出既有效又可靠的新数据收集策略。在实践中，这可能会导致在临床试验中需要更少的患者，或者将确定一种治疗疾病的药物的时间减半。研究人员还计划与高中生和机器学习爱好者进行接触，以提高他们对数据收集的认识水平——例如，即使是一个简单的调查，如果没有精心设计，也可能导致隐私侵犯、人口统计学代表性不足和多种形式的偏见，所有这些都可能导致不准确的结论。对于闭环学习中感兴趣的许多问题，现有技术只关注最小最大最优性，其中最坏情况问题实例的样本复杂性最小化。这种方法导致算法在自然中可能出现的“简单”或良性实例上明显落后，但这些实例远非对抗性。相比之下，本项目将研究在可能近似正确（PAC）设置下交互式学习和强化学习问题的实例最优样本复杂性的基本限制。所获得的见解将应用于算法的设计，自动适应所面临的特定问题实例的内在困难，无论它是良性的还是恶性的。所提出的方法的动机是观察到实例最优样本复杂性分解为一个渐近项，它现在已经很好地表征了，以及一个中等置信度项，已知它在所有实际目的中主导渐近项。由于后一项的性质仍然知之甚少，它的下界将与接近实现它们的算法一起构建。这样的结果将导致算法大大降低总体实例最优样本复杂性，并大大改进倾向于迎合最坏情况的最先进算法。工作将首先集中在结构化线性强盗和强化学习在表格和线性函数近似设置。虽然这些范例在实践中具有广泛的适用性，但它们也具有足够的复杂性，可以将见解外推到更通用的闭环学习范例中。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

项目成果

Beyond No Regret: Instance-Dependent PAC Reinforcement Learning

• 发表时间: 2021-08
• 期刊: ArXiv
• 作者: Andrew J. Wagenmaker;Max Simchowitz;Kevin G. Jamieson

Instance-optimal PAC Algorithms for Contextual Bandits

针对上下文强盗的实例最优 PAC 算法

• 发表时间: 2022
• 期刊: Advances in neural information processing systems
• 作者: Li, Zhaoqi;Ratliff, Lillian;Nassif, Houssam;Jamieson, Kevin;Jain, Lalit
• 通讯作者: Jain, Lalit

Instance-Dependent Near-Optimal Policy Identification in Linear MDPs via Online Experiment Design

• DOI: 10.48550/arxiv.2207.02575
• 发表时间: 2022-07
• 期刊: ArXiv
• 作者: Andrew J. Wagenmaker;Kevin G. Jamieson