CISE: Large: Causal Foundations for Decision Making and Learning

人工智能（AI）在我们的日常生活中无处不在，决策作为一项科学挑战的重要性急剧增加。曾经由人类做出的决定越来越多地委托给自动化系统或在他们的帮助下做出。然而，尽管最近取得了实质性进展，但当前这一代人工智能技术在可解释性、鲁棒性和适应性方面缺乏，这阻碍了人们对人工智能的信任。越来越多的人认识到，强大的决策需要理解环境背后复杂而动态的因果机制，而目前大多数人工智能的形式主义缺乏对因果机制的明确处理。该项目将因果建模和人工智能决策以及学习的力量结合在一起，以产生依赖较少数据的人工智能系统，可以更好地证明和向人们解释他们的决定，更好地对新情况做出反应，因此更安全，更值得信赖。该项目为因果决策系统提供了新的基础——原则、方法和工具。它通过因果成分丰富了传统的人工智能形式化，以实现更高效、更稳健、更通用、更可解释的决策，并有可能从根本上改变人工智能决策领域。该理论将通过机器人和公共卫生中的实际用例进行评估。研究人员将进行广泛的教育工作，并制定培训内容，重点是指导和扩大代表性不足群体的参与。该团队将参与知识转移活动，包括撰写一本关于因果决策的介绍性书籍，以及组织讨论人工智能和决策主题的活动。该项目将结构因果模型框架与人工智能决策的领先方法集成在一起，包括基于模型的马尔可夫决策过程规划及其扩展、强化学习和影响图等图形模型。其结果通过因果建模彻底改变了传统的人工智能决策，朝着开发更高效、稳健、可推广和可解释的决策系统的方向发展。在三个重点中，该项目开发了新的基础（即原则、理论和算法），并为因果授权的决策提供了一个通用的统一框架，该框架概括了主要的决策方法。推力1研究因果决策的基本方面，以保证自主代理和决策支持系统的决策是鲁棒的、样本高效的和精确的。这些目标是通过开发因果关系集成的在线和离线政策学习、干预规划、模仿学习、课程学习、知识转移和适应方法来实现的。推力2研究因果决策的其他方面，这些方面对于人类处于循环中的决策支持系统尤其重要，包括如何利用因果关系来构建解释，决定何时涉及人类，以及赋予系统能力意识和做出符合用户价值观的公平决策的能力。推力3增强了结果工具的可扩展性及其有效推理的能力，在多个目标之间以及在可解释性和决策质量之间进行权衡，并学习世界的因果模型。总之，这些重点将有助于新一代强大的人工智能工具，用于开发自主代理和决策支持系统。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 (AI) has become ubiquitous in our daily lives, and the importance of decision-making as a scientific challenge has increased dramatically. Decisions that were once made by humans are increasingly delegated to automated systems or made with their assistance. However, despite substantial recent progress, the current generation of AI technology is lacking in explainability, robustness, and adaptability capabilities, which hinders trust in AI. There is a growing recognition that robust decision-making requires an understanding of the often complex and dynamic causal mechanisms underlying the environment, while most of the current formalisms in AI lack explicit treatment of causal mechanisms. This project brings together the power of causal modeling and AI decision-making and learning to produce AI systems that rely on less data, can better justify and explain their decisions to people, better react to new circumstances, and consequently are safer and more trustworthy. The project produces new foundations - principles, methods, and tools - for causal decision-making systems. It enriches the traditional AI formalisms with causal ingredients for more efficient, robust, generalizable, and explainable decision-making with the potential to fundamentally transform the AI decision-making field. The theory will be evaluated through real-world use-cases in robotics and public health. The researchers will make extensive educational efforts, and develop training content with a focus on mentorship and broadening the participation of underrepresented groups. The team will engage in knowledge transfer activities including authoring an introductory book on causal decision-making and organizing events to discuss AI and decision-making topics.This project integrates the framework of structural causal models with the leading approaches for decision-making in AI, including model-based planning with Markov decision processes and their extensions, reinforcement learning, and graphical models such as influence diagrams. The outcomes revolutionize traditional AI decision-making with causal modeling toward developing more efficient, robust, generalizable, and explainable decision-making systems. In three thrusts, the project develops new foundations (i.e., principles, theory, and algorithms) and provides a common unified framework for causal-empowered decision-making that generalizes the leading decision-making approaches. Thrust 1 studies essential aspects of causal decision-making to guarantee that the decisions of autonomous agents and decision-support systems are robust, sample-efficient, and precise. These goals are realized by developing methods for causality-integrated online and offline policy learning, interventional planning, imitation learning, curriculum learning, knowledge transfer, and adaptation. Thrust 2 studies additional aspects of causal decision-making that are especially important for decision-support systems where humans are in the loop, including how to exploit causality for constructing explanations, decide when to involve humans, and endow the systems with competence awareness and the ability to make fair decisions that align with the values of their users. Thrust 3 enhances the scalability of the resulting tools and their ability to reason efficiently, trade-off between both multiple objectives and between explainability and decision quality, and learn a causal model of the world. Together, these thrusts will contribute to a new generation of powerful AI tools for developing autonomous agents and decision-support 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.