Actual Causality in a Logical Setting

Alexander BochmanIJCAI 2018 · DOI

TL;DR

This paper proposes a novel logical framework for defining "actual causality" – pinpointing specific causes of particular events. It leverages a causal calculus and a regularity-based approach (NESS test) to offer a more robust definition than many counterfactual methods, particularly for complex scenarios like preemption and overdetermination.

What Problem Does This Paper Address?

Distinguishing between general causation (e.g., "smoking causes cancer") and actual causation (e.g., "John's smoking caused his cancer") is a long-standing challenge in philosophy and AI. While structural equation models (SEMs) have been influential, they often struggle with nuanced cases like preemption (where a cause is overridden) and overdetermination (where multiple causes exist). Many existing counterfactual approaches falter in these scenarios. This paper aims to provide a formal, logical definition of actual causality that can handle these complexities by returning to a regularity-based approach, specifically an enhanced version of the NESS (or INUS) condition, within a formal causal calculus.

Methods

The authors build upon the "causal calculus," a nonmonotonic logical formalism for reasoning about actions and change. They represent causal relationships as rules of the form $A \Rightarrow B$ ("A causes B"). Structural equation models are shown to be representable within this calculus.

The core of their contribution is a new definition of actual causality based on the NESS test, reinterpreted within the causal calculus. Key elements include:

  • Causal Calculus: A logical framework for general causation.
  • Parsimonious Clausal Theories: Causal theories are restricted to clausal rules (literals in the body and head) and must be "parsimonious" (non-redundant). This ensures that causal rules represent distinct mechanisms.
  • Actual Sub-theory: For a given "causal world" (a specific state of affairs), they identify an "actual sub-theory" consisting only of the causal rules that are "active" in that world.
  • NESS Test Formalization: An event $l_0$ is an actual cause of $l$ in a world $a$ if there exists a set of literals $L \subseteq a$ such that $l_0$ and $L$ together causally imply $l$ (using causal inference on the actual sub-theory), and $L$ alone does not imply $l$. This formalizes the "necessary for the sufficiency" aspect of NESS.
  • Regular Inference: They demonstrate that actual causation reasoning can be effectively performed using "regular inference" relations, which are a form of transitive inference, rather than the full causal inference.

Datasets

Not discussed in the paper. The paper focuses on formal definitions and theoretical examples rather than empirical evaluation on datasets.

Key Insights and Ideas

  • Regularity vs. Counterfactuals: The paper argues for the continued relevance and superiority of regularity-based approaches (like NESS) over purely counterfactual ones for actual causality, especially in complex scenarios.
  • Parsimony is Crucial: The requirement for causal theories to be parsimonious (non-redundant) is essential for correctly identifying actual causes and avoiding spurious attributions.
  • Actual Causality is Nonmonotonic: Unlike general causation, actual causation claims can be overridden by adding new causal rules, highlighting its inherently nonmonotonic nature.
  • Logical Formalism for Nuance: The causal calculus provides a powerful and flexible logical framework to precisely capture the subtleties of actual causation, including preemption and overdetermination, which often confound simpler models.
  • Syntax Sensitivity: The definition's sensitivity to the syntactic form of causal rules is presented not as a flaw, but as an inherent feature of actual causation itself, mirroring the original INUS condition.

Key Results

The paper provides a formal definition of actual causality (Definition 8) and establishes key theoretical results:

  • Theorem 1: Actual causal inference within the actual sub-theory ($ \Rightarrow_a $) is equivalent to regular inference ($ \Rightarrow_{reg} $) on that sub-theory.
  • Corollary 2: This recasts actual causation as a direct formalization of the NESS test using regular inference.
  • Corollary 3: Regularly equivalent causal theories support the same actual causation claims.

The paper then demonstrates the effectiveness of their definition through several examples and counterexamples from the literature, including:

  • Loader Example: Correctly identifies C as the cause, not A.
  • Symmetric Overdetermination (Window): Both B and S are correctly identified as causes.
  • Early Preemption (Backup): A is identified as the cause, even though B would have caused the outcome.
  • Bogus Prevention: Correctly identifies ¬A as the cause, not B.
  • Late Preemption (Bottle): ST is identified as a cause, while BT is not.
  • Inevitable Shock: A is identified as the sole actual cause, despite the outcome being inevitable.
  • Purple Flame: Demonstrates that actual causation is not necessarily transitive, correctly identifying P as a cause of PF, and PF as a cause of I, but not P as a cause of I.

Limitations and Future Work

  • Multi-valued Variables: The current definition is primarily for binary variables. The authors note that the causal calculus has tools to handle multi-valued variables, opening avenues for representing more complex scenarios like trumping preemption.
  • Normality and Defaults: The role of normality and defaults in actual causation reasoning is acknowledged as a significant area for future work, though the causal calculus is stated to have the necessary tools for their formal representation.
  • Comparison with Counterfactual Accounts: A more systematic comparison with counterfactual accounts is planned for future work.
  • Causal Counterfactuals: The study is positioned as part of a larger effort to study causal counterfactuals within the causal calculus.

Summary updated

BibTeX
@inproceedings{bochman2018actual,
  author = {Alexander Bochman},
  title = {{Actual Causality in a Logical Setting}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2018},
  doi = {10.24963/ijcai.2018/239},
}
Figure