Researchers extend causality framework to counterfactual analysis in formal action systems

Computer scientists have developed a counterfactual approach to defining causality within situation calculus—a formal system for reasoning about actions and state changes. The work addresses limitations in how existing frameworks, particularly the influential Halpern-Pearl account of actual causality, handle complex causal explanations in dynamic environments.

Building on prior work#

The foundation for this advance comes from recent efforts to make the Halpern-Pearl framework more expressive. Batusov and Soutchanski previously introduced a notion of achievement cause in situation calculus, enabling reasoners to identify causes of quantified effects across action histories. Their formalism leverages the first-order logic of situation calculus to capture rich descriptions of how agents' actions change the world. However, their work lacked a grounding in counterfactual analysis—the "what if" reasoning that intuition suggests should underlie causal claims.

The counterfactual approach#

The new research proposes defining causality explicitly through counterfactual reasoning: a factor is a cause of an outcome if, absent that factor, the outcome would not have occurred. Within action sequences, this perspective naturally generalizes to achievement causality, allowing systems to explain not just what caused what, but specifically which actions or events brought about particular effects.

Bridging frameworks and open questions#

The authors establish connections between their counterfactual account and Batusov and Soutchanski's earlier approach, clarifying how the two perspectives relate. They also examine how counterfactual reasoning applies to the Halpern-Pearl model, bringing to light subtle issues—especially around how counterfactual logic handles disjunctive effects (outcomes that could result from multiple independent causes). This remains a well-known challenge in causal formalism.

The work is grounded in formal logic and addresses foundational questions about how artificial reasoners should model causation in structured, multi-step processes.