A0606
Title: Human-AI collaboration with partial feedback
Authors: Ruijiang Gao - University of Texas at Dallas (United States) [presenting]
Abstract: As AI becomes increasingly embedded in real-world applications, human-AI collaboration is a daily reality for many workers across diverse fields. Purely autonomous AI systems are rare in practice, making it essential to develop AI systems that can work seamlessly and effectively with human partners. Despite this, much existing research focuses on human-AI collaboration in settings with idealized, full-information feedback, such as access to complete classification labels. In contrast, many practical scenarios only provide partial feedback, where outcomes are observed only for chosen actions, complicating the assessment of both human and algorithmic performance. This limitation is common in business settings like customer service, loan applications, and healthcare. The recent work is discussed on addressing key challenges where humans can help improve traditional algorithmic policy learning under partial feedback: learning from offline observational data, and online contextual bandit problems, where human decision-makers possess valuable but imperfect domain knowledge about optimal actions. By addressing these obstacles, the proposed research seeks to advance responsible AI deployment, enabling more effective human-AI systems that are robust, ethically aligned, and practical for complex real-world applications.
