Credibility discounting in the theory of approximate reasoning

Ronald R. YagerUAI 1990

TL;DR

This paper introduces a novel framework for "credibility discounting" within fuzzy logic-based approximate reasoning systems. It proposes a method to reduce the impact of less credible information, not by negating it, but by making it less specific, akin to nonmonotonic reasoning. This approach offers a principled way to manage conflicting or uncertain evidence in automated reasoning systems.

What Problem Does This Paper Address?

Automated reasoning systems often need to aggregate and reason with information from various sources, each potentially having a different level of trustworthiness or "credibility." Traditional approaches might treat zero credibility as a negation of the information, which can be problematic. For instance, if an agent claims "it is raining," and we assign zero credibility to that agent, concluding "it is not raining" is too strong. A more appropriate response is to acknowledge a lack of knowledge about the weather. This paper addresses this gap by proposing a method to discount information based on its credibility, where zero credibility leads to nullification rather than negation. Furthermore, it explores how credibility itself can be context-dependent, influenced by other, more established pieces of evidence, leading to a form of relative credibility.

Methods

The paper proposes a representational scheme for credibility qualification within Zadeh's theory of approximate reasoning (AR). The core idea is to introduce a "credibility value" ($\alpha$) associated with a proposition. When a proposition "V is A" has credibility $\alpha$, it is discounted to a new proposition "V is B," where the fuzzy set B is derived from A using a discounting function $D(\alpha, A(x))$.

Initially, $\alpha$ is treated as a precise value in the unit interval [0,1]. The paper then extends this to handle linguistic values for credibility (e.g., "high," "low") by representing them as fuzzy subsets of the unit interval. This leads to a "second-order fuzzy set" for the discounted proposition.

A key contribution is the concept of "relative credibility." Here, the credibility $\alpha$ of a proposition is not an intrinsic property but is determined by its compatibility with a set of "preeminent knowledge" (other established propositions). This compatibility is measured using possibility theory, specifically the degree of intersection between the proposition's fuzzy set and the preeminent knowledge. This allows credibility to be context-dependent, aligning with nonmonotonic reasoning principles where conclusions can be retracted in light of new, more reliable information.

The paper also explores alternative interpretations of credibility discounting, including a probabilistic approach that leads to Dempster-Shafer belief structures and a model where credibility is a function of the specific value of the variable being reasoned about.

Datasets

Not discussed in the paper. The paper focuses on theoretical framework development and uses illustrative examples rather than empirical evaluation on specific datasets.

Key Insights and Ideas

  • Credibility vs. Negation: Zero credibility doesn't mean the opposite is true; it means we have no information. This is a crucial distinction for robust reasoning.
  • Discounting as Generalization: Credibility discounting effectively makes information less specific, which is a more appropriate way to handle uncertainty than outright negation.
  • Relative Credibility: The trustworthiness of information can depend on other information. This is a powerful concept for adjudicating conflicts between sources.
  • Nonmonotonicity: The framework naturally aligns with nonmonotonic reasoning, allowing conclusions to be revised as new evidence emerges.
  • Linguistic Credibility: The ability to use linguistic terms like "high" or "low" for credibility makes the system more interpretable and user-friendly.

Key Results

The paper presents theoretical results and illustrative examples rather than quantitative performance metrics. Key results include:

  • A formal definition of credibility discounting in approximate reasoning using fuzzy sets.
  • A method for calculating relative credibility based on the compatibility with preeminent knowledge.
  • Demonstrations of how credibility discounting can reduce conflicts between propositions.
  • Exploration of how linguistic credibility values and context-dependent credibility can be incorporated.
  • Connections to Dempster-Shafer theory and probabilistic discounting.

Limitations and Future Work

The authors acknowledge several areas for future work:

  • Generalization of Discounting Functions: While specific discounting functions (like max or exponential) are discussed, a broader exploration of suitable functions is suggested.
  • Metric-Based Discounting: The paper hints at using metrics on the base set X to define credibility, especially when dealing with specific values (e.g., "V is xo"). This could lead to more nuanced discounting that "softens the edges" rather than creating sharp distinctions.
  • Practical Implementation: While the framework is theoretical, its application in real-world expert systems and decision-making processes would be a valuable next step.
  • Comparison with Other Uncertainty Formalisms: Further comparison and integration with other uncertainty formalisms beyond Dempster-Shafer could be beneficial.

Summary updated

BibTeX
@inproceedings{yager1990credibility,
  author = {Ronald R. Yager},
  title = {{Credibility discounting in the theory of approximate reasoning}},
  booktitle = {Conference on Uncertainty in Artificial Intelligence},
  year = {1990},
}
Figure