The Assumptions Behind Dempster's Rule

Nic WilsonUAI 1993

TL;DR

This paper provides a novel axiomatic justification for Dempster's rule of combining belief functions. By formalizing the concept of a "combination rule" and introducing two natural constraints and two reasonable assumptions, the authors uniquely derive Dempster's rule, offering a fresh perspective on its foundations.

What Problem Does This Paper Address?

Dempster's rule is a cornerstone of Dempster-Shafer Theory, a framework for representing and reasoning with uncertainty. It's used to combine evidence from multiple independent sources. However, the rule has faced criticism, particularly regarding the interpretation and justification of "independence" assumptions. This paper aims to provide a more rigorous and convincing justification for Dempster's rule by developing an axiomatic approach that explicitly states the underlying assumptions.

Methods

The authors introduce a formal framework for "source structures," which represent uncertain information over frames of propositions. They then define a "combination rule" as a mapping that takes multiple source structures and produces a single combined source structure. The core of their approach lies in defining two constraints and two assumptions that a combination rule must satisfy:

  • Constraints:
    • Respecting Contradictions: If a combination of evidence leads to a contradiction, the resulting belief must be zero.
  • Respecting Zero Probabilities: If any individual piece of evidence assigns zero probability to a proposition, the combined evidence must also assign zero probability.
  • Assumptions:
    • Assumption (A) - Bayesian Conditioning Commutes: This assumption states that Bayesian conditioning (updating beliefs based on new certain evidence) should yield the same result whether applied before or after combining multiple pieces of evidence.
  • Assumption (B) - Certain Evidences Don't Conflict: This assumption deals with combining evidence from multiple sources where some sources represent certainties. It posits that these certainties should not alter the beliefs derived from other sources, provided they don't conflict.

The paper then proves that these four conditions uniquely determine Dempster's rule of combination.

Datasets

Not discussed in the paper.

Key Insights and Ideas

  • Formalizing Combination Rules: The paper moves beyond informal discussions by precisely defining what a "combination rule" entails, breaking it down into components like the underlying frame, compatibility function, and probability function.
  • Explicit Assumptions: The axiomatic approach makes the assumptions underlying Dempster's rule explicit, allowing for a clearer understanding of when and why the rule is applicable. This addresses a key criticism of previous justifications.
  • Uniqueness of Dempster's Rule: The theorem demonstrating that the proposed constraints and assumptions uniquely determine Dempster's rule is a significant contribution, providing a strong theoretical foundation.
  • Connection to Bayesian Conditioning: Assumption (A) highlights a crucial link between Dempster-Shafer reasoning and Bayesian inference, suggesting a potential for greater integration between these two paradigms.

Key Results

The paper's main result is a theorem stating that Dempster's rule of combination is the unique C-rule (the probability function component of a combination rule) that satisfies the two constraints (respecting contradictions and zero probabilities) and the two assumptions (Bayesian conditioning commutes and certain evidences don't conflict). This provides a strong axiomatic justification for Dempster's rule.

Limitations and Future Work

The authors acknowledge that the reasonableness of Assumption (A) and (B) in all situations is not fully established. Further work is suggested to clarify the conditions under which these assumptions are indeed appropriate. They also note that if these assumptions don't hold, the proof structure might still be useful for justifying Dempster's rule for specific subclasses of belief functions or source structures where the assumptions are more palatable (e.g., simple support functions, consonant support functions). The paper also briefly touches upon the non-normalized version of Dempster's rule, suggesting that while it might be mathematically simpler, its assumptions are less clear.

Summary updated

BibTeX
@inproceedings{wilson1993assumptions,
  author = {Nic Wilson},
  title = {{The Assumptions Behind Dempster's Rule}},
  booktitle = {Conference on Uncertainty in Artificial Intelligence},
  year = {1993},
}
Figure