Observations on Cognitive Judgments

David A. McAllesterAAAI 1991 · OpenReview

Abstract

It is obvious to anyone familiar with the rules of the game of chess that a king on an empty board can reach every square. It is true, but not obvious, that a knight can reach every square. Why is the first fact obvious but the second fact not? This paper presents an analytic theory of a class of obviousness judgments of this type. Whether or not the specifics of this analysis are correct, it seems that the study of obviousness judgments can be used to construct integrated theories of linguistics, knowledge representation, and inference.

Introduction

The distinction between what is logically true and what is cognitively "obvious" is a fundamental puzzle in both artificial intelligence and linguistics. For instance, a chess player immediately perceives that a king on an empty board can reach any square. However, the fact that a knight can also reach any square—while equally true—is not immediately obvious. This paper explores a formal theory of obviousness judgments, proposing that they can be modeled through a restricted form of inference that avoids the computational "state explosion" common in automated theorem provers.

The Theory of Local Inference

To model the immediacy of human cognitive judgments, we introduce the concept of Local Inference ($\Vdash_R$). Unlike standard logical deduction ($\vdash_R$), which allows for the introduction of arbitrary new formulas to reach a conclusion, local inference is constrained by the Subexpression Constraint. A proof is considered local only if every expression it contains is a subexpression of the premises or the conclusion.

This constraint ensures that "obviousness" is computationally tractable. We demonstrate that for any fixed set of local rules, determining whether a statement is locally derivable can be achieved in polynomial time relative to the length of the statement. This aligns with the psychological reality that obvious judgments are made rapidly and without conscious effort.

Knowledge Representation and Induction

Modeling human reasoning requires a language capable of expressing sets and relations. We utilize a Knowledge Representation (KR) language based on class expressions (e.g., "red nodes"), binary relations (e.g., "neighbor-of"), and the Kleene Star ($R^*$) for transitive closure (e.g., "reachable by any number of steps").

Initial set of inference rules for class inclusion and transitive closure.

The challenge lies in formalizing induction—the process of concluding that a property holds for all steps in a sequence. While standard induction is often non-local because it requires "inventing" an inductive hypothesis, we can define rule sets that treat certain inductive leaps as local relative to the problem's existing terms.

Modeling Inductive Obviousness

To better capture human-like inductive reasoning, we developed a "nearly local" rule set, $I'$. This set introduces the preserves operator, which allows the system to recognize when a property is maintained across a relation without needing to construct complex intermediate steps.

The expanded rule set I' which introduces the 'preserves' operator to model human-like induction.

Using these rules, we can formalize why certain truths feel immediate. For example, the judgment that a rat cannot escape a sealed box is "obvious" because the property of "being inside the box" is preserved by any movement the rat can make. Within the $I'$ framework, such conclusions are derived without the computational complexity of general first-order logic.

Results and Discussion

Our analysis reveals that obviousness is a matter of internal consistency within a restricted logical framework, rather than a direct reflection of external truth. This is exemplified by the "bouncing ball" paradox: while it is cognitively obvious that a ball stops bouncing after a finite time, mathematical models of geometric decay suggest an infinite number of bounces. Our theory predicts the "obvious" (but physically simplified) judgment based on how we represent the problem.

The tractability lemma proves that local inference is an $O(n^k)$ process, providing a theoretical basis for why these judgments are "fast." However, limitations remain. The current KR language is simplified, and a complete theory would require a robust translation layer from natural language to these internal logical forms.

Conclusion

By treating obviousness as a distinct cognitive category governed by local inference rules, we can bridge the gap between formal logic and human psychology. This approach offers a promising path toward AI systems that reason in ways that are more aligned with human intuition and less prone to the intractability of general theorem provers.

Summary updated

Abstract

It is obvious to anyone familiar with the rules of the game of chess that a king on an empty board can reach every square. It is true, but not obvious, that a knight can reach every square. Why is the first fact obvious but the second fact not? This paper presents an analytic theory of a class of obviousness judgments of this type. Whether or not the specifics of this analysis are correct, it seems that the study of obviousness judgments can be used to construct integrated theories of linguistics, knowledge representation, and inference.
BibTeX
@inproceedings{mcallester1991observations,
  author = {David A. McAllester},
  title = {{Observations on Cognitive Judgments}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1991},
  url = {https://openreview.net/forum?id=SyWKP1ZO-B},
}
Figure