CAREER: Optimism in Causal Reasoning via Information-theoretic Methods

对现象的因果进行推理是人工智能发展的一个基本问题。从数据中进行因果推理在从工程、计算机科学到医学研究等多个学科中也发挥着关键作用。在过去的几十年里，Pearl（1995）发展了一种正式的概率因果关系数学理论。在这种形式主义中，已经提出了几个算法，说明在定义良好的假设下，可以从数据中提取多少定性和定量的因果知识。这些算法采用了最坏情况的观点：如果因果问题的答案不是唯一的，它们返回的结果是不可识别的。然而，这种方法不适用于许多在不同程度上违反这些关键假设的现实世界系统。研究者认为，通过识别数据中不太可能偶然发生的简单因果解释，有可能显著扩展因果关系理论的适用性。这个项目将通过对大多数模型进行因果推理，而不是在最坏的情况下，将因果关系理论扩展到更广泛的现实世界实例集。为了扩大最先进的因果推理形式主义的范围，研究者将开发新的算法，从数据中识别信息-理论上对潜在因果系统的简单解释。第一个重点是通过基于因果系统熵的奥卡姆剃刀的信息理论解释，发展从观测数据中学习因果关系的方法。第二个重点将分析信息-理论上简单的解释如何帮助近似计算在最坏情况下无法识别的因果关系。第三个重点将利用前两个重点的结果来开发实验设计算法，以便通过干预有效地学习因果结构和因果效应。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。

Reasoning about the causes and effects of phenomena is a fundamental problem in the development of artificial intelligence. Causal reasoning from data also plays a key role in several disciplines, from engineering and computer science to medical research. A formal mathematical theory of probabilistic causation has been developed in the last few decades by Pearl (1995). Several algorithms that illustrate how much qualitative and quantitative causal knowledge can be extracted from data under well-defined assumptions have been proposed within this formalism. These algorithms employ a worst-case view: if the answer to a causal question is not unique, they return that the result is not identifiable. However, such an approach is unsuitable for many real-world systems that violate these crucial assumptions to varying degrees. The investigator argues that it is possible to significantly expand the applicability of causality theory by identifying simple causal explanations in the data that are unlikely to occur by chance. This project will extend the theory of causation to a much wider set of real-world instances by enabling causal reasoning for most models rather than in the worst case.To expand the scope of the state-of-the-art causal reasoning formalism, the investigator will develop novel algorithms that identify information-theoretically simple explanations of the underlying causal system from data. The first thrust seeks to develop methods to learn causal relations from observational data via an information-theoretic interpretation of Occam’s razor based on the entropy of the causal system. A second thrust will analyze how information-theoretically simple explanations can help approximately compute causal effects that are not identifiable in the worst case. A third thrust will leverage the results of the first two thrusts to develop experimental design algorithms for efficiently learning causal structures and causal effects via interventions.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.