Numerical representations of acceptance

Didier Dubois, Henri PradeUAI 1995

TL;DR

This paper introduces "acceptance functions" as a new way to model belief states, extending beyond traditional probability and belief measures. These functions capture the idea of accepting a proposition when confidence in it strictly exceeds confidence in its negation, leading to a deductively closed and consistent set of accepted beliefs. This framework offers a more general setting for belief revision, particularly by exploring conditioning.

What Problem Does This Paper Address?

The paper addresses the challenge of formally modeling belief revision, a core problem in artificial intelligence and philosophy of logic. Existing frameworks, like the AGM postulates, rely on specific notions of uncertainty (e.g., necessity measures) and epistemic entrenchment. The authors question whether other uncertainty measures could support belief revision and explore a more general approach. They propose that a belief state can be directly characterized by a confidence measure that satisfies specific properties related to "acceptance," defined as having strictly greater confidence in a proposition than its negation. This leads to a need for a new class of confidence measures, termed "acceptance functions," that ensure the accepted beliefs form a deductively closed and consistent set.

Methods

The core of the paper is the formal definition and characterization of "acceptance functions." These are monotonic set functions (confidence measures) that satisfy two key properties:

  1. Closure under logical consequence: If proposition A is accepted and A entails B, then B must also be accepted. This is a standard property of confidence measures due to monotonicity.
  2. Closure under conjunction: If propositions A and B are accepted, then their conjunction A ∧ B must also be accepted. This is a novel and crucial requirement that constrains the nature of the confidence measure.

The authors then establish general properties of these acceptance functions, notably proving that they are characterized by a "kernel subset" K. This kernel defines the set of propositions that are "most strongly" accepted. They also explore the duality of acceptance functions and investigate which existing uncertainty measures (belief functions, probability measures, necessity/possibility measures) qualify as acceptance functions. Finally, they introduce the concept of conditioning for acceptance functions, aiming to align it with belief revision.

Datasets

Not discussed in the paper.

Key Insights and Ideas

  • Acceptance as a distinct concept: The paper argues that "acceptance" (confidence in A > confidence in ¬A) is a distinct and useful concept for modeling belief, going beyond mere degrees of belief.
  • The crucial role of conjunction closure: The requirement that the conjunction of accepted propositions must also be accepted is a strong constraint. It implies that not all standard uncertainty measures (like general probability or belief functions) can directly define an "acceptance" belief set.
  • Kernel characterization: Acceptance functions are elegantly characterized by a kernel subset K. A proposition A is accepted if and only if it "contains" K in a specific sense, providing a clear structural insight.
  • Duality and indifference level: Acceptance functions have a dual counterpart, and the "indifference level" (where confidence in A equals confidence in ¬A) plays a role in their characterization.
  • Belief functions and probability measures as special cases: While not all belief functions or probability measures are acceptance functions, the paper identifies specific conditions under which they are. This highlights that acceptance functions are a more general class.
  • Conditioning for belief revision: The paper proposes a way to define conditioning for acceptance functions, aiming to provide a more general foundation for belief revision than existing methods.

Key Results

  • Acceptance functions are characterized by a kernel subset K, such that any proposition A containing K is accepted, and propositions neither containing nor being contained by K have equal confidence in their affirmation and negation.
  • Necessity measures and possibility measures are identified as acceptance functions.
  • Belief functions are acceptance functions if they have a "core subset" (a focal subset contained in all others) or specific singleton focal subset properties.
  • Probability measures are acceptance functions only under very specific conditions: either a single outcome has probability > 1/2, or two outcomes have probability = 1/2.
  • The paper introduces the concept of "context-tolerant acceptance functions" which satisfy a stronger conditioning property, but their full characterization remains an open problem.

Limitations and Future Work

  • Context-tolerant acceptance functions: The full characterization of acceptance functions that remain acceptance functions when restricted to any subset (context-tolerant) is an open problem.
  • Belief revision compatibility: While the proposed conditioning for acceptance functions is compatible with some aspects of belief revision, its full compatibility with all AGM postulates needs further investigation. The paper suggests that the "closure under conjunction" property is lost in general conditioning, which is a significant limitation for direct application to revision.
  • Pathological cases: The paper notes that acceptance functions on very small sets (e.g., 1- or 2-element sets) are trivially acceptance functions, and the focus is on more complex scenarios.
  • Practical application: While the framework is general, its practical utility in complex belief revision scenarios beyond possibility measures requires further exploration.

Summary updated

BibTeX
@inproceedings{dubois1995numerical,
  author = {Didier Dubois and Henri Prade},
  title = {{Numerical representations of acceptance}},
  booktitle = {Conference on Uncertainty in Artificial Intelligence},
  year = {1995},
}
Figure