Probabilistic Belief Change: Expansion, Conditioning and Constraining

Frans VoorbraakUAI 1999

TL;DR

This paper argues that "conditioning" in probabilistic belief change is fundamentally different from "expansion" in classical belief revision. Instead, the authors propose "constraining" as the more appropriate probabilistic analogue of expansion, as it better captures the reduction of ignorance while preserving intermediate results.

What Problem Does This Paper Address?

The standard framework for rational belief change, known as the AGM (Alchourrón, Gärdenfors, and Makinson) theory, primarily deals with belief states represented as logically closed sets of propositions. This framework defines three core operations: expansion (adding consistent information), revision (incorporating inconsistent information), and contraction (removing information). While this has been successful, it often leads to an abstract view of belief change, particularly when moving to probabilistic representations of belief.

In probabilistic settings, "conditioning" (updating probabilities based on new evidence) is often seen as the direct probabilistic counterpart to expansion. However, this paper questions this equivalence. The authors observe that classical expansion is about adding knowledge and reducing ignorance, while probabilistic conditioning seems more focused on reducing uncertainty. They aim to clarify the relationship between these operations and propose a more fitting probabilistic analogue for expansion.

Methods

The paper explores different models of belief states:

  • Belief Sets: Logically closed, consistent sets of propositions.
  • Probabilistic Models: Probability functions on propositions.
  • Partial Probabilistic Models: Classes of probability functions, which can represent both ignorance and uncertainty.

The authors then analyze two key operations on these models:

  • Conditioning: Updating a probability function P with a new piece of information φ, resulting in P(ψ|φ).
  • Constraining: Updating a partial probabilistic model Π with a new piece of information φ, resulting in a subset of Π where φ holds with certainty (Π&φ = {P ∈ Π : P(φ) = 1}).

They compare these operations by examining how they interact with belief sets and how they affect "ignorance" (lack of knowledge about probabilities) versus "uncertainty" (spread of probability values). They also introduce generalized versions of conditioning and constraining that operate on probabilistic languages (LPROB).

Datasets

Not discussed in the paper.

Key Insights and Ideas

  • Conditioning vs. Expansion: The core insight is that probabilistic conditioning and classical belief expansion are fundamentally different. Expansion is about adding knowledge and reducing ignorance, while conditioning is primarily about reducing uncertainty.
  • Constraining as Probabilistic Expansion: The paper argues that "constraining" is a better candidate for probabilistic expansion. Constraining reduces ignorance and allows for a preservative process of building up belief states from ignorance, mirroring classical expansion.
  • Ignorance vs. Uncertainty: The distinction is crucial. Belief sets represent ignorance (what is known/unknown), while probability functions represent uncertainty (degrees of belief). Partial probabilistic models can represent both. Constraining primarily reduces ignorance, while conditioning primarily reduces uncertainty.
  • Preservative Change: Classical expansion allows for learning new information without discarding old beliefs, a "preservative" process. Constraining exhibits this property in probabilistic settings, whereas conditioning does not always.
  • Specific vs. General Information: The paper suggests that "specific" information (about a particular case) should be handled by conditioning, while "general" information (about prior probabilities) calls for constraining.

Key Results

  • Proposition 1: Both constraining and extended conditioning are compatible with the expansion of belief sets when moving from a belief set to its corresponding partial probabilistic model.
  • Proposition 2 & Corollary 3: Generalized constraining translates into the expansion of probabilistic belief sets in LPROB.
  • Example 1 & 4: These examples demonstrate that conditioning (even generalized forms) does not always preserve intermediate results, unlike constraining.
  • Example 5 & 6: These examples illustrate the distinction between specific and general information, arguing that Peter's alibi (specific) should be handled by conditioning, while information about Big Boss's decision process (general) might be handled by constraining.

Limitations and Future Work

  • Provisional Measures: The authors mention that their proposed measures for uncertainty and ignorance in partial probabilistic models are provisional and require further justification.
  • Axiomatic Frameworks: While they discuss axiomatic approaches to belief change (e.g., Dubois and Prade, Grove and Halpern), the paper focuses on conceptual arguments rather than a full axiomatic development of constraining as expansion.
  • Distinction between Specific/General Information: The paper acknowledges that the distinction between specific and general information can be hard to make precise in practice, though they offer a convincing argument for Peter's alibi being specific.
  • Future Work: The paper implicitly suggests further exploration of the axiomatic properties of constraining and conditioning, and a deeper investigation into measures of ignorance and uncertainty. They also note that their discussion of revision is limited to expansion.

Summary updated

BibTeX
@inproceedings{voorbraak1999probabilistic,
  author = {Frans Voorbraak},
  title = {{Probabilistic Belief Change: Expansion, Conditioning and Constraining}},
  booktitle = {Conference on Uncertainty in Artificial Intelligence},
  year = {1999},
}

Loading related content...

Figure