TL;DR
This paper bridges the gap between logical and numerical approaches to belief change, showing how Bayesian conditioning and Lewis' imaging have direct counterparts in possibility theory. It introduces new results for possibilistic conditioning and imaging, demonstrating their adherence to established postulates, and proposes a novel "focusing" operation.
What Problem Does This Paper Address?
Representing and updating knowledge in artificial intelligence often involves dealing with uncertainty. Two prominent frameworks exist: logical formalisms (like those based on AGM postulates for belief revision) and numerical formalisms (like probability theory). While both aim to model how agents change their beliefs when new information arrives, they often operate in different paradigms. This paper addresses the challenge of unifying these perspectives, specifically by exploring how established numerical operations in probability theory (Bayesian conditioning and Lewis' imaging) can be translated into the realm of possibility theory. This translation is crucial for developing more robust and coherent systems for reasoning under uncertainty, especially when dealing with imprecise or qualitative information. The paper also aims to clarify the distinction between "revision" (updating beliefs to be consistent with new information) and "updating" (modifying beliefs to reflect a change in the world state).
Methods
The paper systematically compares and contrasts belief change mechanisms in probabilistic and possibilistic frameworks.
- Probabilistic Framework: It reviews Bayesian conditioning (using Bayes' theorem) and Lewis' imaging (which involves shifting probability mass to "closest" worlds). It highlights justifications for Bayesian conditioning, including information-theoretic and axiomatic approaches.
- Possibilistic Framework: The core contribution lies in defining and analyzing the possibilistic counterparts of conditioning and imaging.
- Possibilistic Conditioning: Two main rules are presented: an ordinal rule (using min) and a product-based rule (similar to Dempster's rule). The paper demonstrates that these rules satisfy key postulates for belief revision, analogous to the AGM framework.
- Possibilistic Imaging: This operation is defined by mapping worlds to their closest neighbors that satisfy the new information, mirroring Lewis' imaging. The paper shows that this possibilistic imaging satisfies the updating postulates of Katsuno and Mendelzon.
- "Focusing" Operation: A novel operation is introduced, distinct from revision and imaging, which can be seen as changing the reference class rather than revising the body of knowledge.
- Uncertain Inputs: The paper extends these frameworks to handle uncertain information, such as "the probability of A is $\alpha$" or "the necessity of A is $\alpha$," by proposing rules analogous to Jeffrey's rule for probabilities.
Datasets
Not discussed in the paper. The paper focuses on theoretical formalisms and illustrative examples rather than empirical evaluation on specific datasets.
Key Insights and Ideas
- Deep Analogy: The paper reveals a strong, fundamental analogy between Bayesian conditioning and Lewis' imaging in probability theory and their corresponding operations in possibility theory. This suggests a unified underlying structure for belief change across different numerical uncertainty formalisms.
- Possibilistic Framework's Completeness: The possibilistic framework, particularly through its conditioning and imaging rules, appears to more naturally accommodate the postulates of both belief revision (AGM) and updating (Katsuno and Mendelzon) than the probabilistic framework alone.
- Distinction Between Revision and Updating: The paper clearly delineates the conceptual difference between revising beliefs to be consistent with new facts (e.g., AGM) and updating beliefs to reflect a change in the world state (e.g., Katsuno and Mendelzon). Possibilistic imaging aligns well with the latter.
- "Focusing" as a New Perspective: The introduction of the "focusing" operation offers a distinct way to handle belief change, emphasizing a shift in perspective or reference class rather than a direct revision of knowledge. This could be particularly useful in scenarios where the underlying world state might not have changed, but our way of observing or interpreting it has.
- Handling Uncertain Information: The proposed methods for handling uncertain inputs (e.g., "necessity of A is $\alpha$") provide a principled way to integrate less certain information into belief change processes, bridging the gap between precise logical inputs and probabilistic representations.
Key Results
- Possibilistic conditioning rules (both ordinal and product-based) satisfy the key postulates for belief revision.
- Possibilistic imaging rules satisfy the key postulates for updating.
- The paper provides formal definitions and justifications for these possibilistic operations, drawing parallels with probabilistic counterparts.
- A novel "focusing" operation is introduced for possibility theory.
- Methods for handling uncertain inputs in possibility theory are presented, analogous to Jeffrey's rule in probability.
Limitations and Future Work
- Logical Framework Integration: While the paper bridges numerical formalisms, it acknowledges that more work is needed to fully integrate these probabilistic and possibilistic rules with syntactic belief change in logical frameworks.
- Syntactic Level of Change: The paper notes that while probabilistic representations of epistemic states have been thoroughly studied for change, the syntactic level (i.e., changing logical formulas themselves) remains less explored.
- Evidence Theory Justification: The justification of different rules in evidence theory is still in its infancy and requires further investigation.
- "Focusing" in Logical Settings: The authors suggest that the "focusing" idea, as opposed to revising a body of knowledge, might be worth introducing in logical settings as well.
- Coherence Between Approaches: Pushing further the consequences of coherence between numerical and symbolic approaches to knowledge representation is identified as a challenging but important future task.
Summary updated