Inference with Possibilistic Evidence

Fengming Song, Ping LiangUAI 1993

TL;DR

This paper introduces "possibilistic evidence" as a way to bridge the gap between fuzzy concepts and probabilistic reasoning, especially when dealing with expert knowledge expressed in natural language. They develop a unified framework for inference that maintains both the compatibility of fuzzy concepts and the consistency of probabilistic logic, allowing for a deeper understanding of how these two types of uncertainty interact.

What Problem Does This Paper Address?

In real-world scenarios, expert knowledge about random events is often expressed using vague, natural language terms (e.g., "very accurate"). This type of information is inherently fuzzy. However, traditional probability theory struggles to directly incorporate such fuzzy descriptions. While Dempster-Shafer's theory of evidence can handle uncertainty arising from a lack of statistics, it doesn't directly address the fuzziness of concepts used by experts. This paper aims to reconcile these two forms of uncertainty by proposing a method to extract probabilistic information from fuzzy, natural language propositions, thereby enabling more robust inference in situations where expert knowledge is imprecise.

Methods

The core of the paper is the development of a unified framework for inference with "possibilistic evidence." This concept is defined as a possibility distribution (a fuzzy subset) that also functions as a body of evidence (a basic probability assignment). Key methodological contributions include:

  • Extending Possibilistic Evidence: The authors extend the concept of possibilistic evidence to handle infinite universes of discourse, incorporating the measurability of set-valued mappings.
  • Unified Inference Framework: They propose a framework that integrates fuzzy logic (for concept compatibility) and probabilistic logic (for reasoning about randomness). This framework allows for the systematic investigation of various fuzzy operators and probability relations.
  • Falling Shadow/Coverage Function: The paper utilizes the concept of a "falling shadow" or "coverage function" of random subsets to connect fuzzy subsets with probabilistic representations. This allows a fuzzy subset to be represented by a random subset, bridging the gap between fuzziness and probability.
  • Possibilistic Evidence from Fuzzy Subsets: A crucial result is the demonstration that a fuzzy subset can induce a unique possibilistic evidence (a basic probability assignment) that extracts the maximum possible information about randomness from the fuzzy proposition.
  • Operations on Possibilistic Evidence: The paper defines and analyzes operations like union, intersection, negation, and conditioning for possibilistic evidence, extending these concepts to handle fuzzy inputs. These operations are shown to have interpretations related to synonymous and antonymous concepts.

Datasets

Not discussed in the paper. The paper focuses on theoretical framework development and does not present empirical evaluations on specific datasets.

Key Insights and Ideas

  • Fuzzy Concepts Contain Probabilistic Information: The paper argues that fuzzy concepts used in natural language descriptions of random events implicitly contain probabilistic information. The challenge is to extract this information systematically.
  • Bridging Fuzziness and Probability: The concept of "possibilistic evidence" and the use of falling shadow functions provide a concrete mechanism to translate fuzzy descriptions into probabilistic terms, enabling unified inference.
  • Synonymy and Antonymy in Inference: The framework allows for the analysis of how the relationship between fuzzy concepts (synonymous or antonymous) affects inference results, providing bounds on plausibility and belief measures.
  • Unified Inference Operators: Various fuzzy inference operators (like Zadeh's union/intersection, Lukasiewicz) can be unified within this framework by selecting appropriate combined probability measures, demonstrating a deep connection between fuzzy and probabilistic inference.
  • Measurability as a Key Concept: The measurability of fuzzy subsets and set-valued mappings is crucial for ensuring the mathematical rigor of the proposed framework.

Key Results

The paper establishes theoretical results that:

  • Define possibilistic evidence for infinite universes of discourse.
  • Show that a fuzzy subset F can be represented by an equivalent subclass of random subsets, and a specific random subset (ξ₀) within this subclass represents the "fundamental theorem of possibilistic evidence."
  • Demonstrate how operations like union and intersection on fuzzy concepts translate to operations on their corresponding possibilistic evidence.
  • Provide inequalities for plausibility, belief, and commonality functions under synonymous and antonymous concept relationships for both union and conditioning operations.
  • Show that common fuzzy inference operators can be derived within the proposed framework by choosing specific combined probability measures.

Limitations and Future Work

  • Computational Complexity: The paper acknowledges that while the framework provides theoretical unification, direct application of probability inference, especially using Dempster's Rule of Combination in infinite universes, can be computationally prohibitive. The authors suggest that using specific probability measures (P₁₂ and P₁₂) might offer bounds for faster inference.
  • Determining Concept Relationships: Accurately determining whether concepts are synonymous or antonymous in implication is noted as a difficult task.
  • Generalizing to More Than Two Sources: The current work primarily focuses on two sources of knowledge. Extending the framework to handle more than two bodies of evidence is suggested for future work.
  • Measuring Randomness Information Value: The paper raises the question of how to best measure the "randomness information value" of inference results, suggesting the difference between Plausibility and Belief as a potential metric, but noting its limitations for intersection operations.
  • Practical Applications: While the framework is theoretical, future work could explore its application in specific domains where expert knowledge is crucial and inherently fuzzy.

Summary updated

BibTeX
@inproceedings{song1993inference,
  author = {Fengming Song and Ping Liang},
  title = {{Inference with Possibilistic Evidence}},
  booktitle = {Conference on Uncertainty in Artificial Intelligence},
  year = {1993},
}
Figure