Nonmonotonic Modes of Inference

Victor JaureguiAAAI 2008

Abstract

This paper introduces a formal framework for "majority-based" default reasoning using modal logic. Unlike traditional nonmonotonic logics that mimic the structure of deductive inference, this "majority" mode reveals that intuitive default reasoning often breaks standard logical rules, such as the principle that the intersection of two likely truths is itself a likely truth.

Introduction

In classical deductive logic, inference is "monotonic": adding new information never invalidates previously held conclusions. However, human reasoning is inherently nonmonotonic—we make plausible guesses based on what is "usually" true and revise them as new facts emerge. While researchers have developed many systems for this (such as Default Logic or Circumscription), most treat nonmonotonic inference as a "normal" modality, assuming it obeys structural rules similar to mathematical proofs. Victor Jauregui argues that these models fail to capture majority-based inference, where something is inferred if it holds in "most" possible scenarios. This paper provides a formal, modal-logic characterization of this specific, non-standard mode of inference.

Formal Framework and Semantics

The author employs Modal (Conditional) Logic to construct a uniform framework for majority reasoning.

  1. Possible-Worlds Semantics: The logic is grounded in a set of possible worlds ($W$). A formula $\phi$ is inferred by default if it holds in a "majority" subset of these worlds.
  2. Defining "Majority": A formal definition of "majority" is established: it must be closed under supersets, and the complement of a majority cannot itself be a majority.
  3. The $\Lambda$ Logic: A new logic, $\Lambda$, is proposed, defined by specific axioms and rules.
Possible-worlds semantics where a majority set $S$ supports the inference of formula $\phi$.

Key Insights: The Failure of Normality

The most significant finding is that majority-based reasoning is "non-normal." It breaks several rules taken for granted in other nonmonotonic systems:

  • The "AND" Problem: In most logics, if you infer $A$ and you infer $B$, you must infer $A \land B$. In majority logic, this fails. If $A$ is true in 70% of worlds and $B$ is true in 70%, their intersection could be as low as 40%—no longer a majority.
  • Failure of Modus Ponens: One might believe $\phi$ is usually true and that $\phi$ usually implies $\psi$, yet $\psi$ itself may only hold in a minority of cases.
  • Beyond Preference: Traditional AI models use "preferential" semantics (focusing on "best" or "most preferred" worlds). Jauregui demonstrates that majority reasoning is fundamentally different and cannot be reduced to these traditional "normal" boxes.

Results and Axiomatization

The paper provides Soundness and Completeness theorems, showing that the algebraic axioms of $\Lambda$ perfectly match the intended majority semantics. It also proves the Finite Model Property, ensuring the logic is decidable (computable).

Standard inference rules. The paper identifies which of these (like RCEA and RCM) apply to majority logic and which (like AND) do not.

The minimal set of axioms ($ID, ABS, CDD$) and rules ($RCEA, RCM$) defining majority-based inference distinguishes it from "normal" systems like those proposed by Kraus, Lehmann, and Magidor. Despite these differences, the system remains "supraclassical," preserving all conclusions of classical logic while sanctioning additional default inferences.

Conclusion and Future Work

Jauregui's work clarifies the logical landscape by identifying majority logic as a distinct, non-normal mode of inference. While the current model focuses on finite sets to avoid paradoxes of infinite majorities, future work involves extending this to infinite possible-world spaces and exploring the context-dependency of "majority" definitions in practical AI agents.

Reference: Jauregui, V. (2008). Nonmonotonic Modes of Inference. Proceedings of the Twenty-Third AAAI Conference on Artificial Intelligence.

Summary updated

BibTeX
@inproceedings{jauregui2008nonmonotonic,
  author = {Victor Jauregui},
  title = {{Nonmonotonic Modes of Inference}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2008},
}
Figure