Collaborative Research: CIF: Medium: Statistical and Algorithmic Foundations of Efficient Reinforcement Learning

合作研究：CIF：媒介：高效强化学习的统计和算法基础

• 批准号: 2221009
• 负责人: Yuxin Chen
• 金额: $ 40万
• 依托单位: University of Pennsylvania
• 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: 2022-01-01 至 2025-09-30
• 项目状态: 未结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=2221009&HistoricalAwards=false
• 关键词: Collaborative; Research; CIF; Medium; Statistical

As a data-driven paradigm for sequential decision making in unknown environments, Reinforcement Learning (RL) has received significant interest in recent years owing to its potential ability to solve difficult problems associated with future societal and scientific developments. However, the explosion of both model dimensionality and complexity in current and emerging applications exacerbates the challenge of achieving efficient RL in sample-starved situations, where data collection is expensive, time-consuming, or even high-stake, e.g., in clinical trials, online advertising, and autonomous systems. As a result, understanding and improving the sample and computational efficiencies of RL algorithms, sometimes under additional resource and system-level constraints, are rightly understood as critical to the successful deployment of RL in the future. In this project the PIs are involving students at all levels with diverse backgrounds in Electrical and Computer Engineering, and in Statistics, are developing education modules on RL to enrich the curriculum, and are co-organizing workshops and outreach activities to enable the broader dissemination of the project outcomes.Despite decades-long research efforts, the statistical and computational underpinnings of RL are still far from being well understood, especially when it comes to finite-sample and finite-time issues which are of crucial operational value. This research project is bridging the theory-practice gap of modern algorithmic approaches to RL. It is doing so by (i) characterizing fundamental limits for the sample and computations complexities in various RL settings, (ii) by developing performance guarantees and uncertainty quantification schemes, and (iii) by designing new computationally efficient algorithms that are provably near-optimal in terms of sample complexity in both single-agent and multi-agent settings. The expected outcomes will enable the trustworthy adoption of RL algorithms in sample-starved environments. The complementary expertise of the research team is being leveraged to enrich the statistical and algorithmic foundations of RL through model-, policy-, and value-based approaches. New efficient algorithms that rely on function approximation schemes are being developed in order to address the curse of dimensionality; the resulting techniques are intended to lead to non-asymptotic analysis tools that deal with the complicated statistical dependencies present in RL. This rich research agenda is expected to foster multidisciplinary efforts at the intersection of high-dimensional statistics, non-convex optimization, control theory, information theory, and machine learning.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.

作为未知环境中顺序决策的数据驱动范式，强化学习（RL）由于其解决与未来社会和科学发展相关的难题的潜在能力，近年来引起了极大的兴趣。然而，当前和新兴应用中模型维度和复杂性的爆炸式增长加剧了在样本匮乏的情况下实现高效强化学习的挑战，在这种情况下，数据收集昂贵、耗时，甚至风险很高，例如在临床试验中，在线广告和自治系统。因此，理解和提高 RL 算法的样本和计算效率（有时在额外的资源和系统级限制下）被正确地理解为对于未来成功部署 RL 的关键。在这个项目中，PI 吸引了电气和计算机工程以及统计学领域具有不同背景的各级学生参与，正在开发 RL 教育模块以丰富课程，并共同组织研讨会和外展活动，以更广泛地传播 RL 知识。尽管经过数十年的研究努力，强化学习的统计和计算基础仍然远未得到充分理解，特别是在涉及具有至关重要的操作价值的有限样本和有限时间问题时。该研究项目正在弥合现代强化学习算法的理论与实践差距。它通过以下方式实现这一目标：(i) 表征各种 RL 设置中样本和计算复杂性的基本限制，(ii) 开发性能保证和不确定性量化方案，以及 (iii) 设计可证明接近最优的新计算高效算法就单代理和多代理设置中的样本复杂性而言。预期结果将使强化学习算法在样本匮乏的环境中得到可靠的采用。研究团队的互补专业知识正在通过基于模型、策略和价值的方法来丰富强化学习的统计和算法基础。为了解决维数灾难，正在开发依赖函数逼近方案的新高效算法；由此产生的技术旨在产生非渐近分析工具，用于处理强化学习中存在的复杂统计依赖性。这一丰富的研究议程预计将促进高维统计、非凸优化、控制理论、信息论和机器学习交叉领域的多学科努力。该奖项反映了 NSF 的法定使命，并通过评估被认为值得支持基金会的智力价值和更广泛的影响审查标准。

项目成果

Fast Global Convergence of Natural Policy Gradient Methods with Entropy Regularization

具有熵正则化的自然策略梯度方法的快速全局收敛

• DOI: 10.1287/opre.2021.2151
• 发表时间: 2022-01
• 期刊: Operations Research
• 影响因子: 2.7
• 作者: Cen, Shicong;Cheng, Chen;Chen, Yuxin;Wei, Yuting;Chi, Yuejie
• 通讯作者: Chi, Yuejie

Sample Complexity of Asynchronous Q-Learning: Sharper Analysis and Variance Reduction

异步 Q-Learning 的样本复杂性：更清晰的分析和方差减少

• DOI: 10.1109/tit.2021.3120096
• 发表时间: 2022-01
• 期刊: IEEE Transactions on Information Theory
• 影响因子: 2.5
• 作者: Li, Gen;Wei, Yuting;Chi, Yuejie;Gu, Yuantao;Chen, Yuxin
• 通讯作者: Chen, Yuxin

Policy Mirror Descent for Regularized Reinforcement Learning: A Generalized Framework with Linear Convergence

正则化强化学习的策略镜像下降：具有线性收敛的广义框架

• DOI: 10.1137/21m1456789
• 发表时间: 2023-06
• 期刊: SIAM Journal on Optimization
• 影响因子: 3.1
• 作者: Zhan, Wenhao;Cen, Shicong;Huang, Baihe;Chen, Yuxin;Lee, Jason D.;Chi, Yuejie
• 通讯作者: Chi, Yuejie

Breaking the sample complexity barrier to regret-optimal model-free reinforcement learning

打破样本复杂性障碍，实现后悔最优无模型强化学习

• DOI: 10.1093/imaiai/iaac034
• 发表时间: 2023-02
• 期刊: Information and Inference: A Journal of the IMA
• 作者: Li, Gen;Shi, Laixi;Chen, Yuxin;Chi, Yuejie
• 通讯作者: Chi, Yuejie

Pessimistic Q-Learning for Offline Reinforcement Learning: Towards Optimal Sample Complexity

用于离线强化学习的悲观 Q-Learning：实现最佳样本复杂度

• 发表时间: 2022-01
• 期刊: International Conference on Machine Learning
• 作者: Laixi Shi;Gen Li;Yuting Wei;Yuxin Chen;Yuejie Chi
• 通讯作者: Yuejie Chi