Possibilistic Conditioning and Propagation

Yen-Teh HsiaUAI 1994

TL;DR

This paper provides a logic-based axiomatization for "confidence transfer," a method for updating beliefs in possibility theory. It argues that Dempster's rule of conditioning is the most natural choice for this update, as it uniquely supports the notion of belief independence, a crucial concept for reasoning. The work then shows how this rule enables efficient, local computation of beliefs in complex systems, akin to Bayesian networks.

What Problem Does This Paper Address?

Possibility theory offers a way to formalize beliefs that are not necessarily precise, unlike probability theory. A key challenge in possibility theory is how to update beliefs when new information arrives – a process known as conditioning. Several "confidence transfer" rules have been proposed, but there's no clear consensus on which one is theoretically sound or how they relate to fundamental principles of belief. This paper aims to provide a rigorous justification for confidence transfer and identify the most appropriate rule by grounding it in logical inference and the intuitive concept of belief independence.

Methods

The authors build upon the work of Gärdenfors and Makinson on expectation-based inference. They strengthen Rott's proposal for belief revision by introducing a new axiom (D1) and deriving a set of axioms (D1-D3) that characterize a consonant belief function. This axiomatization leads to the concept of "confidence transfer" as a general algorithm for updating possibility measures.

To select the "best" confidence transfer rule, they introduce the notion of belief independence. Two propositions are considered belief independent if learning about one doesn't affect one's belief in the other. They then test existing conditioning rules (Yager's, Dubois and Prade's minimum rule, and Dempster's rule) against this independence criterion.

Finally, they demonstrate how Dempster's rule, once established as the preferred method, can be used with a "causal net" technique for local computation of conditional marginal possibilities. This is analogous to how Bayesian networks perform efficient reasoning in probabilistic settings.

Datasets

Not discussed in the paper. The paper uses theoretical examples and a well-known "alarm problem" from Pearl [88] for illustration.

Key Insights and Ideas

  • Logic-Based Justification for Confidence Transfer: The paper provides a formal, axiomatic derivation of confidence transfer, moving beyond ad-hoc definitions.
  • Belief Independence as a Guiding Principle: The concept of belief independence is presented as a crucial, intuitive criterion for evaluating different conditioning rules.
  • Dempster's Rule as the "Natural" Choice: Dempster's rule is uniquely identified as the only rule among those considered that satisfies the belief independence criterion. This is a significant finding, as it provides a strong theoretical basis for its use.
  • Analogy to Bayesian Networks: The paper highlights how Dempster's rule enables a local computation scheme for possibility theory, mirroring the efficiency of Bayesian networks in probability theory. This suggests a deep structural similarity in how uncertainty can be managed in these different frameworks.
  • Distinction from Probability: While the formal structure of possibility theory might resemble probability theory, the paper emphasizes that their fundamental notions of belief are opposite, preventing a simple transformation.

Key Results

  • An axiomatization of confidence transfer is derived from logical principles and expectation-based inference.
  • Dempster's rule of conditioning is shown to be the only rule among the considered ones that supports belief independence.
  • A local computation scheme for possibility theory, analogous to Bayesian network inference, is presented, leveraging Dempster's rule. This scheme allows for efficient computation of conditional marginal possibilities in complex systems.

Limitations and Future Work

  • Ruling out All Alternatives: The paper acknowledges that while Dempster's rule is the "only natural choice" and the only one that supports belief independence among the tested rules, it cannot definitively rule out all infinitely many possible normalization rules for confidence transfer.
  • Intuitive Legitimacy: The authors anticipate questions about why belief independence was chosen as the criterion and whether the probabilistic flavor of their independence definition hints at possibility theory being a mere transformation of probability theory. They defend their choice by arguing for the a priori nature of belief independence and emphasize that structural similarity does not imply equivalence.
  • Further Theoretical Validation: The paper suggests that it remains to be shown that Dempster's rule is the only variation of confidence transfer that supports belief independence (in the sense of Theorem 3.1).

Summary updated

BibTeX
@inproceedings{hsia1994possibilistic,
  author = {Yen-Teh Hsia},
  title = {{Possibilistic Conditioning and Propagation}},
  booktitle = {Conference on Uncertainty in Artificial Intelligence},
  year = {1994},
}

Loading related content...

Figure