CIF: Medium: Foundations of Learning from Paired Comparisons and Direct Queries

直接查询和配对比较是各种科学学科中流行的数据获取手段。这两种类型的数据已经被分开研究，然而，几种现代应用--尤其是涉及人的判断的应用--需要分析这两种类型的数据的组合。这种分析为随机建模、实验设计和算法开发带来了新的机遇和挑战。这项研究涉及建立从直接测量和配对比较中学习的统一观点，目的是了解各种感兴趣问题的基本限制和权衡，并开发和实施可以利用这两种类型数据的实用算法。这些算法的实用性将通过在材料科学和众包等不同领域的应用来展示。本项目专注于函数估计、特征选择和优化等具体技术问题，这些技术问题可以利用被动和主动获得的直接查询和配对比较，目标如下。第一个目标是从不同统计模型下的配对测量和直接测量的组合中建立学习的基本信息论界限。这些基本限制还涉及了解两种测量形式之间的内在权衡，这两种形式的测量解释了不同的噪音和成本。第二个目标是开发可扩展和计算效率高的算法，其性能达到这些基本限制。第三个目标是利用直接的实验和成对的专家反馈，将理论转化为应用，如材料设计。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

Direct queries and paired comparisons are popular means of data acquisition in various scientific disciplines. These two types of data have been studied separately, however, several modern applications -- particularly those involving human judgments -- require analysis of a combination of these two types of data. Such an analysis presents novel opportunities and challenges for stochastic modeling, experimental design and algorithm development. This research involves establishing a unified view of learning from direct measurements and paired comparisons with the aim of understanding fundamental limits and tradeoffs for various problems of interest, and developing and implementing practical algorithms that can leverage both types of data. The utility of the algorithms will be demonstrated by application to diverse domains including material science and crowdsourcing.This project focuses on the specific technical problems of function estimation, feature selection, and optimization that can leverage passively and actively acquired direct queries and paired comparisons, with the following objectives. The first objective is to establish fundamental information-theoretic limits of learning from a combination of paired and direct measurements under various statistical models. These fundamental limits also involve understanding the inherent tradeoffs between the two forms of measurements that accounts for the different noise and costs. The second objective is to develop scalable and computationally efficient algorithms whose performance attains these fundamental limits. The third objective is to transfer the theory to practice in applications such as material design using direct experimental and pairwise expert feedback.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.