On implementing usual values

Ronald R. YagerUAI 1986

TL;DR

This paper introduces a formal system for reasoning with "usual" values, a common but often implicit aspect of human commonsense knowledge. By combining fuzzy sets and probability theory into "Possibility-Probability granules," the authors provide a mechanism to represent and manipulate these imprecise, default-like pieces of information, paving the way for more robust AI systems.

What Problem Does This Paper Address?

Artificial intelligence systems often struggle with commonsense reasoning, which relies heavily on understanding what is "usual" or "typical" in a given situation. For example, knowing that "a cup of coffee usually costs fifty cents" is more accurate than a strict logical statement. Current non-monotonic logic systems don't fully capture this nuanced, probabilistic, and somewhat vague nature of "usualness." The authors aim to bridge this gap by developing a formal framework that can represent and reason with these "usual values" in a computationally tractable way.

Methods

The core of the paper is the introduction and development of Possibility-Probability (Poss-Prob) granules. These are data structures designed to represent knowledge about a variable's value that is both probabilistic and "granular" (imprecise).

  • Canonical Statements: A basic form of knowledge is a "canonical statement" like "V is A," where V is a variable and A is a fuzzy subset of possible values. This signifies that V's value lies within A, representing possibilistic uncertainty.
  • Poss-Prob Granules: A more general structure, a Poss-Prob granule, associates a set of canonical statements (V is A_k) with probabilities (p_k). This means there's a probability p_k that the statement "V is A_k" is the correct restriction on V. The probabilities p_k sum to one.
  • Fuzzy Focal Elements: Unlike some related structures (like Dempster-Shafer belief structures), the "focal elements" (A_k) in Poss-Prob granules can be fuzzy subsets, allowing for greater expressiveness.
  • Measures of Plausibility and Belief: The paper extends the concepts of plausibility and belief measures to work with fuzzy focal elements, providing bounds on the probability of events.
  • Translation Rules: The authors develop logical translation rules to convert linguistic statements involving "usually" into these Poss-Prob granule structures. This includes handling conditional statements ("if...then") and compound logical operations ("and," "or").
  • Arithmetic Operations: The paper also outlines how to perform arithmetic operations (like addition, subtraction) on these usual values, extending the calculus beyond purely logical manipulations.

Datasets

Not discussed in the paper.

Key Insights and Ideas

  • "Usuality" as a Blend of Probability and Granularity: The paper's central thesis is that "usual" knowledge is inherently a combination of probabilistic uncertainty (it's not always true) and granularity (the "usual" value itself is often imprecise, best represented by a fuzzy set).
  • Poss-Prob Granules as a Unified Representation: These granules offer a powerful way to unify different types of uncertainty and knowledge representation, particularly for commonsense reasoning.
  • Formalizing "Usually": The paper provides a concrete mathematical interpretation for linguistic statements like "usually V is A," defining it as a Poss-Prob granule where the probability of V being in A is "at least alpha" (where alpha is close to 1).
  • Bridging Logic and Arithmetic: The development of both logical and arithmetic operations on these "usual values" is crucial for building practical AI systems that can reason about the world in a more human-like way.

Key Results

The paper presents a formal framework and a set of rules for:

  • Representing "usual" statements as Possibility-Probability granules.
  • Translating compound linguistic statements involving "usually" into these formal structures.
  • Performing logical operations (conjunction, disjunction) on these granules.
  • Performing arithmetic operations on variables qualified by "usual" values.
  • Demonstrating that "usually qualified" propositions provide tighter bounds on probabilities compared to "usually squared qualified" propositions.

Limitations and Future Work

  • Computational Complexity: While the framework is powerful, the paper doesn't delve deeply into the computational complexity of these operations, which could be a significant factor in practical implementations.
  • Real-World Data: The paper focuses on the theoretical framework and doesn't present empirical evaluations on real-world datasets.
  • Further Development of Arithmetic: The paper outlines the approach to arithmetic operations but suggests that further work is needed to fully develop this calculus, particularly for fuzzy arithmetic.
  • Integration with Existing Systems: Future work could focus on integrating this system into existing knowledge-based systems and expert systems to leverage its capabilities for commonsense reasoning.

Summary updated

BibTeX
@inproceedings{yager1986implementing,
  author = {Ronald R. Yager},
  title = {{On implementing usual values}},
  booktitle = {Conference on Uncertainty in Artificial Intelligence},
  year = {1986},
}
Figure