Embracing Causality in Formal Reasoning

Judea PearlAAAI 1987

Introduction

In the development of artificial intelligence, traditional logic has often struggled to replicate human-like common sense. A primary reason for this failure, as argued by Judea Pearl, is that standard logic treats all "if-then" rules as identical material implications. It ignores the fundamental distinction between a rule describing a cause (e.g., Fire $\rightarrow$ Smoke) and one describing evidence (e.g., Smoke $\rightarrow$ Fire). By failing to account for the directionality of inference, purely syntactic systems often fall into traps of circular reasoning or fail to correctly "explain away" facts when new information arrives.

The Problem: The Joe Paradox

To illustrate the limitations of default logic, Pearl presents the "Joe" example. Consider two facts and a rule:

  • Fact A: Joe is over 7 years old.
  • Fact B: Joe can read and write.
  • Rule: If Joe can read and write ($B$), he is likely over 7 ($A$).

In a standard logical system, any evidence supporting $B$ would automatically support $A$. However, human reasoning distinguishes between how $B$ was established.

The Joe example: Case 1 shows how evidential support propagates, while Case 2 illustrates how causal expectations should block certain inferences.

In Case 1, we learn Joe recites Shakespeare ($e_1$). This is evidence he can read ($B$), which in turn provides evidence he is over 7 ($A$). This is a valid evidential chain. In Case 2, we learn Joe’s father is an English professor ($e_2$). While this makes it likely Joe can read ($B$), it does not necessarily imply he is over 7; he might just be a precocious 5-year-old. In fact, the father's profession "explains away" the reading ability, making us less certain about the age requirement than we would be if we didn't know about the father.

The C-E System: A Causal Framework

Pearl proposes the C-E System, a coarse logical abstraction designed to prevent these "un-human" logical leaps. The system categorizes rules and beliefs into two types:

  1. Labeling Rules: Every default rule is labeled as either C-def (Causal/Expectation-evoking, denoted $\xrightarrow{c}$) or E-def (Evidential/Explanation-evoking, denoted $\xrightarrow{e}$).
  2. Causal: Fire $\xrightarrow{c}$ Smoke (Fire causes smoke).
  3. Evidential: Smoke $\xrightarrow{e}$ Fire (Smoke is evidence of fire).
  4. Labeling Beliefs: Propositions are labeled based on their support: C-believed (supported by expectation) or E-believed (supported by evidence).

Inference Propagation Rules

The system defines how these labels propagate through a network:

  • (a) Causal rule + Causal belief $\rightarrow$ Causal belief.
  • (b) Causal rule + Evidential belief $\rightarrow$ Causal belief.
  • (c) Evidential rule + Evidential belief $\rightarrow$ Evidential belief.

Crucially, the system excludes the propagation of an evidential rule from a causal belief. This prevents a predicted consequence (like Joe reading because of his father) from being used as evidence for a hidden cause (Joe's age).

Key Insights: Explaining Away and Asymmetry

A central feature of this framework is its ability to handle "explaining away." In causal reasoning, if multiple causes can produce the same effect, finding evidence for one cause should weaken our belief in the others.

The wet grass scenario: A classic causal network where a state (Q) can be explained by multiple causes (P1, P2) or lead to various effects (R1, R2).

In the "wet grass" scenario shown above, if we observe that the grass is wet ($Q$), we might suspect it rained ($P_1$). However, if we then discover the sprinkler was on ($P_2$), our belief in $P_1$ should decrease. The C-E system achieves this by treating causal rules as "censors." Once a belief is established via a causal path (expectation), it is blocked from triggering further evidential inferences that would lead to competing causes.

Conclusion and Limitations

The C-E System successfully resolves paradoxes that stumped previous default logics by acknowledging the asymmetry of human reasoning. It demonstrates that even a non-numeric, categorical distinction between cause and effect can significantly improve AI reasoning.

However, the system has limitations. It is strictly binary and cannot handle varying degrees of "strength" in expectations or evidence—a task Pearl notes is better suited for full numerical Bayesian analysis. Additionally, while it identifies when to be skeptical, it does not yet provide a complete mechanism for retracting beliefs once they are definitively disproven. Future work suggests a "rough quantization" of rule strength to bridge the gap between pure logic and probability.

Summary updated

BibTeX
@inproceedings{pearl1987embracing,
  author = {Judea Pearl},
  title = {{Embracing Causality in Formal Reasoning}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1987},
}
Figure