A Critique of Yoav Shoham's Theory of Causal Reasoning

Antony GaltonAAAI 1991

In his 1991 AAAI paper, Antony Galton presents a fundamental challenge to Yoav Shoham’s then-influential theory of causal reasoning. Galton argues that Shoham’s framework mistakenly incorporates human knowledge and "changeable" logic into the very definition of causality, rather than treating causality as an objective reality of the physical world. This critique remains a pivotal moment in the history of AI, marking a shift from purely epistemic models toward those that respect the ontological distinctions between events and states.

The Problem: Conflating Physics and Psychology

In the late 1980s, the AI community was deeply invested in solving the "frame problem"—the challenge of representing what remains unchanged when an action is performed. Yoav Shoham proposed a popular theory that defined causal laws using modal logic (specifically knowledge and belief) and the concept of "chronological ignorance." This approach allowed for non-monotonic reasoning, where an agent can retract a conclusion when new information appears.

Galton identifies a major conceptual gap in this approach: Shoham’s theory conflates the physical laws of the world with the mental state of the agent reasoning about them. If we define a cause based on what we "know" or "believe," we risk suggesting that the laws of physics change simply because an agent learns something new. Galton argues that this "subjectivist" approach leads to logical paradoxes and fails to capture the true nature of cause and effect.

Objectivism vs. Subjectivism

Galton introduces a rigorous distinction between two philosophical stances:

  1. Objectivism: The view that causality exists "out there" in the world, independent of any observer. Causal laws are universal and monotonic.
  2. Subjectivism: The view that causality is a feature of our epistemic perspective—a way of organizing our limited knowledge of the world.

Galton argues that while our reasoning about the world is non-monotonic (we change our minds), the causal laws themselves are not. For example, the law "pressing the button rings the bell" should be considered true or false based on the physical wiring, regardless of whether an agent knows the battery is dead. Shoham’s logic, by contrast, makes the law itself dependent on the agent's ignorance.

The Logical Critique

Galton employs logical analysis to test the robustness of Shoham’s formal definitions, specifically focusing on the modal operator $\Box$ (usually interpreted as "necessarily" or "it is known that"). He demonstrates that if this operator follows standard modal logic (such as S5), it leads to a paralyzing conclusion: we can never truly "know" a causal effect will occur because we can never know every possible background condition.

Furthermore, Galton points out a "sleight of hand" in Shoham’s famous example of a car starting. Shoham claims that adding a "connected battery" to the "key turn" cause makes the reasoning non-monotonic. Galton counters that the battery being connected is a state (a background condition), while the key turn is the event (the trigger). By failing to distinguish these "aspectual" categories, Shoham is forced to bake non-monotonicity into the causal law itself.

Proposed Formal Structure

Rather than defining causality through the lens of an agent's knowledge, Galton suggests a revised form for causal rules that separates events from states:

$$\phi_1 \wedge \dots \wedge \phi_m \wedge \chi_1 \wedge \dots \wedge \chi_n \rightarrow \psi$$

In this structure:

  • $\phi$ represents events (discrete occurrences like turning a key).
  • $\chi$ represents states (persisting conditions like a battery being charged).
  • $\psi$ is the resulting effect.

By removing modal operators from the core law, Galton proposes that causal laws should be treated as monotonic and objective. The "non-monotonicity" should be handled by the inference engine that uses those laws, not by the definition of causality itself.

Conclusion and Legacy

Galton’s critique successfully argued for a move away from purely epistemic causal models toward "aspectual" models. While Galton acknowledged that his critique did not offer a complete alternative formal theory in this specific paper, it pointed toward the necessity of "reifying tokens" (specific instances of events) rather than just types. This direction was later explored extensively in the development of Situation Calculus and Event Calculus. The paper remains a classic reminder that in AI, the way we represent the world must distinguish between the reality of the environment and the limitations of the agent's mind.

Summary updated

BibTeX
@inproceedings{galton1991critique,
  author = {Antony Galton},
  title = {{A Critique of Yoav Shoham's Theory of Causal Reasoning}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1991},
}
Figure