Learning Decision Rules in Shifting Environments

该研究项目将为环境变化下的数据驱动决策开发新的方法和软件。数据驱动型决策的大多数工作从根本上依赖于统计环境的稳定性。用于数据驱动的决策制定的可用工具通常假设将部署决策规则的未来环境类似于过去收集数据的环境。然而，由于各种因素，许多实际应用程序显示出显著的环境变化。新开发的方法将扩大研究人员将数据驱动决策应用于环境变化系统的能力。可能的应用领域包括从医疗环境和社会项目到在线市场。教育活动将包括培训研究生和开发部分基于这项研究结果的学习资源。所有的研究产品都将通过公开可用的存储库进行传播，软件将在开源许可下发布。该研究项目将提供一个实用的基于深度学习的框架，用于学习对未知分布变化具有鲁棒性的决策规则。该项目将开发在分布变化未知的情况下学习决策规则的方法，在这些情况下，根据不可观察的特征，一些子群体可能是采样不足或过度的。例如，考虑一项基于志愿者的研究，研究抗抑郁药对抑郁症患者的影响，在研究人群中，采取措施对抗抑郁症的动机可能是一个未被观察到的属性，在研究人群中被过度代表——因此，我们能够收集数据的人群可能与整个患者群体在一些重要但不可观察的属性上有所不同。该项目还将研究在分配变化下的学习决策规则，这种规则是通过社会环境中相互作用的代理（例如，在市场上相互买卖商品）的均衡行为自然产生的，并提供解决由此产生的挑战的方法。总的来说，该项目将提供新的方法和软件解决方案，以及相关的教育资源，这将扩大数据驱动决策可以成功部署的问题类别。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。

This research project will develop new methods and software for data-driven decision making under environmental shift. Most work in data-driven decision making fundamentally relies on the stability of the statistical environment. Available tools for data-driven decision making generally assume that future environments in which decision rules will be deployed resemble the past environment where data was collected. Many real-world applications, however, display significant environmental shift due to various factors. The newly developed methods will expand researchers' ability to apply data-driven decision making to systems with environmental shift. Possible application areas range from medical settings and social programs to online marketplaces. Educational activities will include the training of graduate students and the development of learning resources based in part on the results of this research. All research products will be disseminated via publicly available repositories, and software will be released under an open-source license. This research project will provide a practical deep learning-based framework for learning decision rules that is robust to unknown distributional shifts. The project will develop methods for learning decision rules in settings with unknown distributional shifts, and where some sub-populations may be under- or over-sampled according to unobservable characteristics. Consider, for example, a volunteer-based study on the effects of antidepressants among patients suffering from depression, where motivation to take steps to fight depression could be an unobserved attribute that is overrepresented in the study population – and so the set of people we are able to collect data on may differ from the full patient population along some important but unobservable attributes. The project also will investigate learning decision rules under distributional shifts that naturally arise via equilibrium behavior in social settings with agents who interact with each other (e.g., by buying and selling goods to each other in a marketplace) and provide methods to address the resulting challenges. Overall, this project will provide new methodological and software solutions – as well as associated educational resources – that will expand the class of problems where data-driven decision making can be successfully deployed.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.