EAGER: AI-DCL: Hybrid human algorithm predictions: balancing effort, accuracy, and perceived autonomy

人工智能算法正在成为日常决策场景中越来越常见的组成部分。这些算法使用医疗数据来帮助医生做出诊断和治疗决策，使用传感器数据来辅助驾驶员操作车辆，并使用评论数据来推荐购买的产品和访问的餐馆。这些系统通常是混合的，其决策受到人工智能算法和人类决策者的共同影响。例如，在一种操作模式中，人类在考虑了算法的输入后做出每一个决定。在另一种操作模式中，当人力资源有限时（如在医疗诊断中），混合系统可以使用算法做出大多数常规决策，并只将那些对算法最具挑战性的问题分配给人类决策者。随着这些混合系统越来越多地用于关键的决策任务，解决有关这些系统的理解和设计的问题非常重要。该项目将从算法开发和人类推理和决策的认知模型的双重角度解决关于混合决策系统的一系列基本问题。例如，当考虑到有限的人力资源和人与算法技能之间潜在的权衡时，如何优化混合人-算法系统的整体性能？从心理学的角度来看，缺乏自主性是否会对人类的参与产生负面影响？是否可以设计系统来减轻这种影响？该项目将为人类和算法如何协同工作开发新的理论和方法，其结果将有助于在医疗、交通、商业和消费者应用等各个领域产生更准确、更强大的决策系统。为了实现其目标，该研究项目将汇集心理学、机器学习和贝叶斯估计方面的先前工作线索。该项目将由两个紧密耦合的组件组成，共同的重点是建模和理解预测问题，由人类和算法专业知识的结合来处理。第一个组件将为算法仲裁者开发和评估不同的计算和统计框架，以平衡大规模分类任务中的黑箱预测和人类专业知识。第二个部分将建立在人类认知和心理学的理论基础上，分析联合算法-人类预测性能，并明确考虑人类退出和不继续使用算法的影响（例如，由于感知到缺乏自主性）。在该项目期间，将在各种混合预测场景和不同决策分配方法下进行一系列广泛的用户研究，以支持混合算法-人类预测系统的新认知见解和计算方法的开发。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。

Artificial intelligence algorithms are becoming an increasingly common component in everyday decision-making scenarios. Such algorithms use medical data to assist doctors in making diagnostic and treatment decisions, use sensor data to assist drivers in operating vehicles, and use review data to make recommendations for products to purchase and restaurants to visit. These systems are often hybrid, with decisions influenced by a combination of an artificial intelligence algorithm and a human decision-maker. For example, in one mode of operation, a human makes each decision after taking into account input from the algorithm. In another mode of operation, when human resources are limited (as in medical diagnosis), a hybrid system can use the algorithm to make the majority of routine decisions and allocate to a human decision-maker only those problems that are most challenging to the algorithm. As these hybrid systems are being increasingly used in critical decision-making tasks it is important to address questions about the understanding and design of such systems. This project will address a series of fundamental questions about hybrid decision-making systems from the dual perspectives of algorithm development and cognitive models of human reasoning and decision-making. For example, how can the overall performance of a hybrid human-algorithm system be optimized when taking into account limited human resources and potential trade-offs between human and algorithmic skills? From a psychological perspective, does a perceived lack of autonomy negatively affect human engagement and can systems be designed to mitigate this? The project will develop new theories and methods for how humans and algorithms can work together and the results will help produce more accurate and more robust decision-making systems across a variety of areas such as medicine, transportation, business, and consumer applications. To achieve its goals the research project will bring together prior threads of work from psychology, machine learning, and Bayesian estimation. The project will consist of two closely-coupled components with a common focus on modeling and understanding of prediction problems that are handled by a combination of human and algorithmic expertise. The first component will develop and evaluate different computational and statistical frameworks for an algorithmic arbitrator that balances black-box predictions and human expertise in large-scale classification tasks. The second component will build on theories from human cognition and psychology to analyze joint algorithm-human prediction performance, with explicit consideration of the effect of a human dropping out and not continuing to work with the algorithm (e.g., due to a perceived lack of autonomy). An extensive series of user studies, under a variety of hybrid prediction scenarios and different decision allocation methods, will be conducted during the project to support the development of new cognitive insights and computational approaches for hybrid algorithm-human prediction systems.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.