CAREER: Real-Time Stochastic Optimization with Large Structured Strategy Sets and High-Volume Data Streams

这项教师早期职业发展(CALEAR)研究建议支持用于解决大规模实时随机优化问题的建模和算法框架的开发。这项研究将重点放在潜在不确定性的分布未知的情况下。这种环境中的决策者经常面临大量的备选方案，并且必须基于大量数据流实时做出决策。通过利用每个问题的特定结构，本研究将开发将实时决策与近似算法相结合的方法，以解决具有大策略集的复杂随机优化问题。这项研究将通过与行业合作伙伴合作，调查这些算法在基于搜索的广告和供应链管理问题上的应用。此外，研究结果将被整合到互动教学模块或案例研究中。这项拟议的研究涵盖了当前信息驱动环境下的许多问题，包括基于搜索的广告服务中的广告投放、在线零售商的库存管理以及多臂匪徒问题。如果成功，这项研究将产生一个研究和分析这类问题的统一框架，以及一套适用于这些问题的算法。这项研究将建立新的技术来扩展近似算法的性能保证，以反映系统的长期平均性能，并通过利用每个问题中的特殊结构来提供提高收敛速度的新方法。与业界公司广泛合作并整合这些关系以丰富学生的课堂体验将是该项目的重要成果。

This Faculty Early Career Devevelopment (CAREER)research proposes to support the development of a modeling and algorithmic framework for solving large-scale real-time stochastic optimization problems. This research will focus on a setting where distributions of the underlying uncertainty are not known. The decision maker in this environment often faces a large number of alternatives and must make a decision in real-time based on high-volume data streams. By exploiting specific structures of each problem, this research will develop methodologies that combine real-time decision-making with approximation algorithms for solving complex stochastic optimization problems having large strategy sets. The research will investigate applications of these algorithms to problems in search-based advertising and supply chain management by working with industry partners. In addition, the outcome of the research will be integrated into an interactive teaching module or a case study. The proposed research encompasses many problems in the current information-driven environment, including ad placement in search-based advertising services, inventory management for online retailers, and the multi-armed bandit problem. If successful, the research will result in a unifying framework for studying and analyzing this class of problems, along with a suite of algorithms applicable to these problems. The research will establish new techniques for extending the performance guarantee of approximation algorithms to reflect long-run average performance of the system, and provide new methodologies for improving the convergence rates by leveraging special structures within each problem. Extensive collaboration with companies in the industry and the integration of these relationships to enrich the classroom experience for students will be an important outcome of this project.