Jeffrey's rule of conditioning generalized to belief functions

Philippe SmetsUAI 1993

TL;DR

This paper generalizes Jeffrey's rule of conditioning, a method for updating probabilistic beliefs, to the framework of belief functions. This extension allows for more nuanced belief revision when dealing with uncertain information, particularly when the updating evidence pertains to a coarser representation of the world than the original beliefs.

What Problem Does This Paper Address?

In probability theory, Jeffrey's rule provides a way to update a probability measure when new information becomes available. This new information doesn't necessarily pinpoint a single outcome but rather assigns probabilities to a set of possible outcomes. The challenge arises when we want to apply similar reasoning to belief functions, which are used in the Transferable Belief Model (TBM) and other frameworks to represent degrees of belief that are not necessarily probabilities. Specifically, the paper addresses how to generalize Jeffrey's rule when the updating information (represented by a belief function on a subalgebra) is coarser than the original belief function defined on a larger algebra. This is crucial for situations where new evidence provides information about broader categories rather than specific events.

Methods

The authors generalize Jeffrey's rule by defining two key requirements for the updated belief function, denoted as $P_3$ (or $bel_3$ in the belief function context):

  1. Constraint R1 (or C1): The updated belief function must agree with the new information wherever it is defined. In the belief function context, this means if the new information is represented by a belief function $m_2$ on a subalgebra $B$, the updated belief function $m_3$ must satisfy $bel_3(X) = bel_2(X)$ for all $X \in B$.
  2. Constraint R2 (or C2/C3): This constraint generalizes the probabilistic idea that the relative likelihoods of outcomes within a specific event should be preserved. The paper proposes two main ways to generalize this for belief functions, leading to two distinct generalized rules:
  3. Jeffrey Geometric Rule of Conditioning: This rule uses the "geometrical rule of conditioning" for belief functions. It ensures that the ratios of updated beliefs for elements within the same subalgebra atom are consistent with the ratios of beliefs derived from the initial belief function after conditioning on that atom.
  4. Jeffrey-Dempster Rule of Conditioning: This rule uses "Dempster's rule of conditioning" (which is essentially an unnormalized rule). Similar to the geometric rule, it maintains consistency of belief ratios but uses a different underlying conditioning mechanism.

The paper also introduces additional constraints (C3F and C3R) to handle cases where the original algebra elements intersect multiple atoms of the subalgebra, ensuring the updated belief function is properly defined across the entire original algebra.

Datasets

Not discussed in the paper.

Key Insights and Ideas

  • Belief Functions as a More General Framework: The paper highlights that belief functions offer a richer framework than standard probability measures for representing uncertainty, especially when dealing with incomplete or imprecise information.
  • The Importance of Subalgebras: The generalization hinges on the concept of a subalgebra, allowing for updating information that is "coarser" than the original belief space. This is a key distinction from standard probability conditioning.
  • Two Natural Generalizations: The existence of two distinct, yet justifiable, generalizations (geometric and Dempster-Dempster) for Jeffrey's rule in the belief function context is a significant finding. This reflects the inherent flexibility and nuances of belief function theory.
  • Asymmetry of Jeffrey's Rule: The paper emphasizes that Jeffrey's rule, in its generalized form, maintains an asymmetry where the new information ($bel_2$) plays a more dominant role in defining the updated belief ($bel_3$) within its domain. This contrasts with symmetric combination rules like Dempster's rule of combination.

Key Results

The paper presents two generalized rules for conditioning belief functions, inspired by Jeffrey's rule:

  1. Jeffrey Geometric Rule of Conditioning: This rule is derived by imposing constraints C1, C2F, and C3F. It's characterized by the update formula: $m_3(A) = \frac{m_1(A \cap B(A))}{ \sum_{X \in \mathcal{A}(A)} m_1(X \cap B(A)) } m_2(B(A))$ where $B(A)$ is the smallest element in the subalgebra $B$ containing $A$, and $\mathcal{A}(A)$ is the set of elements in the original algebra that map to the same $B(A)$.

  2. Jeffrey-Dempster Rule of Conditioning: This rule is derived by imposing constraints C1, C2R, and C3R. It's characterized by the update formula: $m_3(A) = \frac{m_1(A | B(A))}{ \sum_{X \in \mathcal{A}(A)} m_1(X | B(A)) } m_2(B(A))$ where $|$ denotes conditioning by Dempster's rule.

The paper also shows that in the limiting case where the subalgebra $B$ is the entire algebra $A$, these rules reduce to standard conditioning.

Limitations and Future Work

  • Nature of Atoms: The paper notes that the nature of the atoms of the subalgebra $B$ is important for applying these rules.
  • Comparison with Other Generalizations: The authors discuss and critique other attempts to generalize Jeffrey's rule, arguing that their proposed constraints (particularly C1) better capture the spirit of Jeffrey's original rule.
  • Bayesian Criticisms: The paper touches upon the ongoing debate between Bayesian and Dempster-Shafer/TBM approaches, suggesting that their generalization offers a way to avoid certain Bayesian criticisms by not assuming a probability measure on all possible spaces.
  • Future Work: The paper implies that further exploration of the practical applications and implications of these generalized rules in various domains would be valuable. The authors also mention that their work covers cases where $bel_1$ is a probability function, which is a specific instance of their broader framework.

Summary updated

BibTeX
@inproceedings{smets1993jeffrey,
  author = {Philippe Smets},
  title = {{Jeffrey's rule of conditioning generalized to belief functions}},
  booktitle = {Conference on Uncertainty in Artificial Intelligence},
  year = {1993},
}
Figure