Modeling uncertain and vague knowledge in possibility and evidence theories

Didier Dubois, Henri PradeUAI 1988

TL;DR

This paper argues that traditional probability theory falls short in modeling certain aspects of uncertain knowledge, particularly vagueness and partial ignorance. It advocates for the integration of possibility theory and evidence theory (like Shafer's belief functions) as complementary tools that can enrich AI systems' ability to reason with nuanced, human-like uncertainty.

What Problem Does This Paper Address?

Artificial intelligence systems often need to reason with incomplete or imprecise information. While probability theory is the dominant framework for uncertainty, it struggles to capture certain types of "uncertain knowledge." Specifically, it can't easily represent situations where we lack information about what we don't know (ignorance) versus situations where we know the odds but the outcome is inherently random (randomness). The paper highlights how probability theory forces a binary "true" or "false" interpretation for propositions, making it difficult to model partial ignorance or vagueness in a natural way. For instance, if we're ignorant about two independent events, probability theory might lead to paradoxical conclusions when trying to combine that ignorance. This limitation hinders AI's ability to truly mimic human expert reasoning, which often deals with these subtler forms of uncertainty.

Methods

The paper proposes a framework that integrates probability theory with possibility theory and evidence theory.

  • Possibility Theory: This theory, rooted in fuzzy sets, uses possibility distributions ($\pi$) to represent uncertainty. It allows for degrees of possibility, where a higher value indicates a greater likelihood or compatibility. The paper shows how possibility measures can be derived from fuzzy sets and how they relate to upper probabilities.
  • Evidence Theory (Belief Functions): This framework, particularly Shafer's theory, uses belief functions (Bel) and plausibility functions (Pl) to represent uncertainty. It allows assigning belief to sets of propositions, not just individual ones, which is crucial for modeling ignorance. The paper demonstrates how belief functions can be constructed from bodies of evidence and how they relate to lower and upper probabilities.
  • Integration: The core idea is to view possibility and belief functions as extensions of probability theory, capable of representing aspects that probability alone cannot. They are presented as ways to model "lower" and "upper" probabilities, providing a richer representation of knowledge. The paper also touches upon Cox's axiomatic framework for probability and argues why it doesn't fully encompass belief and possibility functions.

Datasets

Not discussed in the paper. The paper is theoretical and focuses on formal frameworks for representing uncertainty, not empirical evaluation on specific datasets.

Key Insights and Ideas

  • Ignorance vs. Randomness: A critical distinction is made between uncertainty arising from a lack of knowledge (ignorance) and uncertainty arising from inherent randomness. Probability theory struggles to differentiate these, while possibility and belief functions offer better tools.
  • Limitations of Probability: The paper demonstrates how probability theory can lead to paradoxical results when dealing with partial ignorance, especially when combining uncertain propositions. The forced assignment of probabilities to complementary events (P(a) and P(¬a)) is a key issue.
  • Possibility as Upper Probability: Possibility measures are shown to be equivalent to upper probability measures, providing a way to represent the maximum degree to which a proposition could be true.
  • Belief Functions as Lower Probability: Belief functions are presented as lower probability measures, representing the minimum degree to which a proposition is true based on available evidence.
  • Vagueness as Uncertainty about Meaning: The paper adopts Cheeseman's definition of vagueness as uncertainty about meaning, which can be modeled using probability distributions over possible meanings. This connects vagueness to fuzzy sets and random sets.
  • Complementarity, Not Replacement: The authors emphasize that these new theories are not meant to replace probability but to complement it, enriching the descriptive power of uncertainty modeling in AI.

Key Results

The paper doesn't present quantitative results in terms of accuracy or performance metrics on specific tasks. Instead, its key results are theoretical:

  • Formalization of Limitations: It provides formal arguments and examples illustrating the limitations of probability theory in modeling ignorance, imprecision, and vagueness.
  • Integration Framework: It establishes a theoretical framework where possibility measures and belief functions can be understood as upper and lower probability measures, respectively, thus integrating them with the probabilistic paradigm.
  • Connection to Fuzzy Sets: It shows how fuzzy set membership functions can be interpreted within the framework of possibility theory and evidence theory, providing a unified view.
  • Critique of Existing Views: It critiques certain interpretations and claims about these theories, such as Cheeseman's assertion that possibility theory assumes maximal dependence.

Limitations and Future Work

  • Computational Complexity: While the paper presents theoretical frameworks, the practical implementation and computational efficiency of combining these different uncertainty models for complex AI systems are not deeply explored.
  • Axiomatic Foundations: The paper touches upon Cox's axioms but doesn't fully develop the axiomatic foundations for possibility and belief functions in the same way probability theory has been axiomatized.
  • Bulk of Work Needed: The authors explicitly acknowledge that bringing possibility and evidence theories to the same level of development as probability theory requires significant further research. They mention that many results reported are preliminary.
  • Practical Applications: The paper focuses on the theoretical underpinnings. Future work would involve demonstrating how these richer uncertainty models can be effectively applied to real-world AI problems, such as decision support systems and approximate reasoning systems.

Summary updated

BibTeX
@inproceedings{dubois1988modeling,
  author = {Didier Dubois and Henri Prade},
  title = {{Modeling uncertain and vague knowledge in possibility and evidence theories}},
  booktitle = {Conference on Uncertainty in Artificial Intelligence},
  year = {1988},
}
Figure