Belief Revision Operators with Varying Attitudes Towards Initial Beliefs

Adrian Haret, Stefan WoltranIJCAI 2019 · DOI

TL;DR

This paper explores how agents can revise their beliefs in response to new information, moving beyond the standard "conservative" approach. By introducing new axioms that capture different "attitudes" an agent can have towards its existing beliefs, the authors provide a richer framework for understanding belief revision and its connection to rational choice theory.

What Problem Does This Paper Address?

Belief revision is a fundamental concept in artificial intelligence and logic, modeling how an agent updates its knowledge base when presented with new, trusted information. The standard framework, known as AGM (Alchourrón, Gärdenfors, Makinson) or KM (Katsuno and Mendelzon), includes a set of postulates that define rational belief change. A key assumption in these standard frameworks is a "conservative" attitude: if new information is consistent with existing beliefs, the agent simply adds it, preserving as much of the old information as possible. This paper questions whether this conservative stance is the only rational one, proposing that agents might adopt various attitudes towards their initial beliefs, ranging from extreme conservatism to outright dismissal. The challenge is to formalize these varying attitudes and understand their implications for belief revision.

Methods

The authors tackle this problem by focusing on the semantic interpretation of belief revision operators, which can be viewed as defining a preference ordering (a preorder) over possible worlds. The standard KM postulates correspond to a specific type of preorder. The core of this paper's contribution lies in introducing new axioms (R6-R10) that capture different "attitudes" an agent can have towards its initial beliefs. These axioms are then semantically characterized by modifying the properties of the preorder.

Specifically, they introduce:

  • New Axioms (R6-R10): These axioms relax or modify the standard postulates related to how existing beliefs are treated during revision. For instance, R6 and R7 explore different ways of incorporating new information when it's consistent with existing beliefs, moving away from the strict "add and supplement" rule. R8 and R9 focus on the "dual" knowledge base, representing the negation of existing beliefs. R10 deals with how revisions interact with existing beliefs.
  • Semantic Characterization: The paper establishes a correspondence between these new axioms and specific properties of preorders on possible worlds. For example, axiom R6 is linked to a property where existing beliefs are considered among the most plausible worlds, while R7 allows for more flexibility in ranking these worlds.
  • Concrete Operators: To illustrate these concepts, the authors construct specific belief revision operators based on:
  • Distance Metrics: Various ways to measure the "distance" between possible worlds (e.g., drastic distance, Hamming distance).
  • Aggregation Functions: Functions that combine these distances to rank worlds (e.g., minimizing distance, maximizing distance, lexicographic ordering). These functions define the preorders.

Datasets

Not discussed in the paper. The paper uses propositional logic and theoretical models rather than empirical datasets.

Key Insights and Ideas

  • Beyond Conservatism: The most significant insight is that the standard "conservative" approach to belief revision is just one of many possible rational attitudes an agent can adopt towards its existing beliefs. This opens up a richer landscape for modeling belief change.
  • Attitudes as Axioms: The paper demonstrates that different "attitudes" towards initial beliefs can be formally captured by specific logical axioms, providing a precise way to distinguish between various revision strategies.
  • Preorders as the Semantic Core: The connection between belief revision operators and preorders on possible worlds is central. The paper shows how modifying the properties of these preorders directly corresponds to changing the agent's attitude towards its initial beliefs.
  • Link to Rational Choice Theory: The paper highlights the deep connection between belief revision and rational choice theory, suggesting that belief revision operators can be seen as choice functions over sets of interpretations, guided by an agent's preferences.
  • New Operators for Diverse Attitudes: The construction of concrete operators using distance metrics and aggregation functions provides practical examples of how these varying attitudes can be implemented, offering a spectrum from "minimal" to "maximal" plausibility for existing beliefs.

Key Results

  • Axiomatic Characterization: The paper provides axiomatic characterizations for belief revision operators that exhibit varying attitudes towards initial beliefs.
  • Semantic Representation: It establishes that these operators can be represented by specific types of preorders on possible worlds, which are themselves characterized by new semantic properties.
  • Concrete Operator Examples: Several new belief revision operators are introduced and analyzed, demonstrating how different distance metrics and aggregation functions lead to distinct revision behaviors and attitudes towards initial beliefs. For instance, operators like $\mathcal{O}_{H, \text{min}}$ are shown to be highly conservative, while others like $\mathcal{O}_{H, \text{max}}$ push existing beliefs to be less plausible.
  • Analysis of Stability and Indifference: The paper introduces and analyzes properties like "indifference to already held beliefs" (RIDF) and "stability" (RSTB), showing how they relate to the new axioms and operators.

Limitations and Future Work

  • Focus on Propositional Logic: The current work is limited to propositional logic. Extending these ideas to richer logical formalisms (e.g., first-order logic) is a natural next step.
  • Interplay with Other Axioms: While the paper introduces new axioms and analyzes their semantic properties, further work is needed to fully understand their interplay with the entire set of standard belief revision axioms.
  • Semantic Characterizations for All Axioms: The paper suggests that more work is needed to provide complete semantic characterizations for all the newly introduced axioms.
  • Applications of Varying Attitudes: The authors pose the question of what these different attitudes are "good for," suggesting future work in exploring their utility in tasks like learning and continuous refinement of world representations, potentially connecting belief revision to other AI fields.
  • Iterated Revision and Kinetic Consistency: The paper touches upon iterated revision and kinetic consistency, indicating these as areas for further exploration in the context of varying attitudes.

Summary updated

BibTeX
@inproceedings{haret2019belief,
  author = {Adrian Haret and Stefan Woltran},
  title = {{Belief Revision Operators with Varying Attitudes Towards Initial Beliefs}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2019},
  doi = {10.24963/ijcai.2019/239},
}
Figure