Action-theoretic formalization of actual causation and its applications

对实际因果关系的研究涉及在给定的日志中找到导致观察到的结果的行为或事件。因果关系分析在自动推理中起着至关重要的作用，并且在几乎每个领域都有许多应用，包括计算机科学（例如，用于数据库，程序验证，可解释的人工智能等），制造和工程、医学和健康科学（例如，用于诊断目的）以及侵权法（例如，分配责任和责备），仅举几例。例如，如果一架飞机坠毁，分析飞行记录仪捕捉到的动作并识别导致这场灾难的动作是很有用的。自亚里士多德时代以来，哲学家们一直在努力解决这个基本问题，即究竟是什么导致了一种效应，但还没有提出一个足够普遍的适当定义。事实证明，一般来说，真正的因果关系是非常棘手的。 目前的正式方法，实际的因果关系是基于结构方程模型（SEM）。虽然这些模型非常流行，但它们的表现力有限，并且存在各种问题。该研究计划旨在克服实际因果关系形式化所涉及的一些挑战。为此，我和我的学生将发展一个基于行动和变化的正式理论的实际因果关系的综合理论，并研究其潜在的应用。在短期内，我将追求三个主要目标。首先，我将在离散动力系统中定义实际原因。我将确保我的形式化可以支持非线性场景，即那些观察到的事件只是部分有序的场景，并从客观的角度和个体代理的角度来建模因果关系。其次，我将研究混合动态域中的因果关系，其中变化可以涉及离散事件的发生，也可以是时间流动的结果，并由一些连续函数决定。最后，我将研究这个理论的各种应用。特别是，我将适用于能源系统故障的诊断因果关系。此外，我将解决另外两个特别有影响力的间接应用，即使用根本原因分析来解释代理行为，以及多代理系统中的责任和指责的归属。许多人将实际原因的行动理论形式化视为克服当前建议局限性的关键技术。研究计划的结果将有助于发展正式的理论以及软件工具，用于分析各种实际领域的实际因果关系。我的学生将开发因果关系理论，基于这些实现因果引擎，并通过精心设计的定理和实验来评估它们，沿着为工业和学术界做好准备的方式积累技能。

Research on actual causality involves finding in a given log the actions or events that caused an observed effect. Causality analysis plays a crucial role in automated reasoning and has numerous applications in practically every field, including computer science (e.g. for databases, program verification, explainable artificial intelligence, etc.), manufacturing and engineering, medicine and health science (e.g. for diagnostic purposes), and tort law (e.g. to assign responsibility and blame), to name a few. For instance, if an aircraft crashes, it is useful to analyze the actions captured by the flight-recorder and identify those that led to this disaster. Philosophers since the time of Aristotle have been grappling with this basic question of what actually caused an effect, but a proper definition that is general enough is yet to be proposed. It turns out that actual causality in general is extremely tricky to formulate. Current formal approaches to actual causation are based on Structural Equations Models (SEMs). Although very popular, these models have limited expressiveness and suffer from a variety of problems. This research program aims at overcoming some of the challenges involved in the formalization of actual causation. To this end, my students and I will develop a comprehensive theory of actual causation that is based on a formal theory of action and change, and investigate its potential applications. In the short term, I will pursue three main objectives. First, I will develop a definition of actual cause within discrete dynamical systems. I will ensure that my formalization can support non-linear scenarios, i.e. those where the observed events are only partially ordered, and model causation from both an objective perspective and from the perspective of individual agents. Secondly, I will study causality in hybrid dynamic domains where change can involve discrete event occurrences as well as can be a result of the flow of time and be dictated by some continuous function. Finally, I will investigate various applications of this theory. In particular, I will apply causation for the diagnosis of faults in energy systems. Moreover, I will tackle two other particularly impactful indirect applications, namely the explanation of agent behaviour using root cause analysis, and the attribution of responsibility and blame in multi-agent systems. Action theoretic formalizations of actual cause are seen by many as a key technology for overcoming the limitations of current proposals. The outcome of the research program will contribute to the development of formal theories as well as software tools for analyzing actual causation in a variety of practical domains. My students will develop causality theories, implement causal engines based on these, and evaluate them through carefully designed theorems and experiments, accumulating skills along the way that will prepare them well for industry and academia.