A structural model interpretation of Wright's NESS test

Abstract

Although understanding causation is an essential part of nearly every problem domain, it has resisted formal treatment in the languages of logic, probability, and even statistics. Autonomous artificially intelligent agents need to be able to reason about cause and effect. One approach is to provide the agent with formal, computational notions of causality that enable the agent to deduce cause and effect relationships from observations. During the 1990s, formal notions of causality were pursued within the AI community by many researchers, notably by Judea Pearl. Pearl developed the formal language of structural models for reasoning about causation. Among the problems he addressed in this formalism was a problem common to both AI and law, the attribution of causal responsibility or actual causation. Pearl and then Halpern and Pearl developed formal definitions of actual causation in the language of structural models. Within the law, the traditional test for attributing causal responsibility is the counterfactual "but-for" test, which asks whether, but for the defendant's wrongful act, the injury complained of would have occurred. This definition conforms to common intuitions regarding causation in most cases, but gives non-intuitive results in more complex situations where two or more potential causes are present. To handle such situations, Richard Wright defined the NESS Test. Pearl claims that the structural language is an appropriate language to capture the intuitions that motivate the NESS test. While Pearl's structural language is adequate to formalize the NESS test, a recent result of Hopkins and Pearl shows that the Halpern and Pearl definition fails to do so, and this thesis develops an alternative structural definition to formalize the NESS test.