A Circumscriptive Theory for Causal and Evidential Support

Eunok PaekAAAI 1990

Introduction

In the realm of artificial intelligence and commonsense reasoning, humans naturally distinguish between causal directions (e.g., fire causing smoke) and evidential directions (e.g., smoke serving as evidence for fire). However, formalizing this intuition in logical systems has historically proven difficult. This paper introduces a formal framework using circumscription to model how we reason about causes and evidence, specifically addressing the phenomenon of "causal asymmetry."

The Problem of Causal Asymmetry

A central challenge in AI logic is correctly modeling how different pieces of information interact. Consider two distinct scenarios:

  1. Common Cause (Symmetry): If measles causes both a rash and a fever, observing a rash increases the likelihood of a fever. The effects support each other because they point to a shared underlying cause.
  2. Common Effect (Asymmetry/Competition): If a rash can be caused by either measles or an allergy, these causes "compete." If we observe a rash and then confirm the patient has measles, the likelihood of an allergy actually decreases. The measles "explains away" the rash, removing the evidential need for the allergy.

Early qualitative systems, most notably Judea Pearl’s 1988 framework, often struggled with these nuances. These systems were prone to "all-or-nothing" conclusions, sometimes erroneously concluding that multiple potential causes were all true simultaneously rather than treating them as disjunctive possibilities. This paper seeks to provide a more flexible, justification-based logic that captures these behaviors without the computational overhead of full probability theory.

Methodology: Circumscription and Justification

The proposed framework utilizes Circumscription, a nonmonotonic logic technique developed by John McCarthy. Circumscription allows a system to make "common sense" assumptions by minimizing specific predicates—effectively assuming that an event does not occur unless there is an explicit reason for it.

Key Components

  • Predicate Reification: The system treats causal relationships as formal entities, using structures like causes(Rain, WetGrass) to allow the logic to reason about the relationship itself.
  • The Minimization Axiom: The author introduces a central axiom stating that a property holds if and only if it was triggered by a known cause or occurred "acausally" (spontaneously). By circumscribing these acausal events, the system is forced to find the most parsimonious causal explanation for any given observation.
  • Justification-Based Support: Drawing from Truth Maintenance Systems (TMS), the paper defines "support" not as absolute truth, but as a shift in the justification for a belief. If new information $A$ simplifies or strengthens the logical justification for $B$, then $A$ is said to support $B$.

Key Insights

Explaining Away

The system naturally captures the "explaining away" effect. In the classic "Wet Grass" toy problem, if the grass is wet, both Rain and a Sprinkler are possible justifications. However, once Rain is confirmed, the "need" for the Sprinkler to explain the wetness is removed. In this circumscriptive logic, the justification for the Sprinkler decreases upon the confirmation of Rain.

Partiality and Symmetry

Unlike earlier qualitative frameworks that required every component of a complex cause to be present before making an inference, this system allows for partial support. Furthermore, it successfully distinguishes between the mutual support of two effects (evidential symmetry) and the competition between two causes (causal asymmetry).

Qualitative vs. Quantitative Reasoning

The author argues that while Bayesian networks provide powerful probabilistic tools, a qualitative approach is often more intuitive for human-like reasoning and easier to implement in expert systems where precise probabilities are unknown or unnecessary.

Results and Comparison

The paper demonstrates that this circumscriptive approach avoids the counterintuitive errors found in Pearl’s early qualitative models. In the "Wet Grass" example, rather than accepting both Rain and Sprinkler as true, the system correctly identifies them as a disjunction (Rain OR Sprinkler). By providing formal definitions for "Evidential Support" and "Causal Support" based on minimal support sets, the paper offers a rigorous foundation for non-probabilistic causal reasoning.

Limitations and Future Work

Despite its successes, the framework faces several challenges:

  • Axiomatization: The system assumes that causal rules are already provided; it does not address the massive problem of how to properly axiomatize causality from scratch.
  • Granularity: As a qualitative system, it cannot distinguish between a highly likely cause and a marginal one; all valid justifications are treated with equal logical weight.
  • Computational Complexity: Circumscription is computationally expensive, often reaching NP-hard levels of complexity, which may limit the scalability of the system for large-scale knowledge bases.

Summary updated

BibTeX
@inproceedings{paek1990circumscriptive,
  author = {Eunok Paek},
  title = {{A Circumscriptive Theory for Causal and Evidential Support}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1990},
}
Figure