Shielded Base Contraction (Extended Abstract)

Marco Garapa, Eduardo Fermé, Maurício D. Luís ReisIJCAI 2019 · DOI

TL;DR

This paper introduces and axiomatizes "shielded base contractions," a novel approach to belief change that allows agents to retain certain beliefs even when contracting others. By proposing twenty distinct classes of these contractions and analyzing their interrelations, the authors provide a comprehensive framework for understanding flexible belief revision.

What Problem Does This Paper Address?

The standard model for belief change, known as the AGM (Alchourrón, Gärdenfors, Makinson) model, assumes that when an agent contracts a belief, that specific belief is removed from their knowledge base. However, this "success postulate" can be unrealistic. Agents might have reasons to hold onto certain beliefs, even if they are logically related to a belief they are willing to discard. This paper addresses this limitation by exploring "shielded base contractions," which are designed to be more flexible. Instead of strictly removing a contracted belief, these operators can choose to leave the belief base unchanged if the contracted belief is deemed "irretractable." This work extends previous research on shielded contractions for belief sets to the more practical context of belief bases.

Methods

The core of this paper is the formalization and axiomatic characterization of various "shielded base contractions." The authors define a shielded contraction as an operator that, given a belief base and a sentence to contract, either leaves the base unchanged (if the sentence is "irretractable") or performs a standard contraction (if the sentence is "retractable"). They then explore different ways to define "retractability" by considering various properties for sets of retractable sentences.

The paper systematically defines and analyzes twenty classes of shielded base contractions, each induced by a specific type of base contraction (e.g., partial meet, kernel, basic AGM-generated) and a specific set of properties for retractable sentences. They provide axiomatic characterizations for these classes, meaning they define a set of postulates that precisely capture the behavior of each class. A significant part of the work involves analyzing the interrelations between these classes, specifically investigating which classes are strict subclasses of others. This is visualized through detailed diagrams.

Datasets

Not discussed in the paper.

Key Insights and Ideas

  • Flexibility in Belief Revision: The most significant insight is the introduction of "shielded base contractions" as a more realistic model for belief change. This allows for agents to maintain a more nuanced stance on their beliefs, not strictly adhering to the success postulate of traditional AGM.
  • A Rich Taxonomy of Contractions: By proposing twenty distinct classes of shielded contractions, the paper offers a comprehensive taxonomy. This allows for fine-grained modeling of different agent behaviors and their willingness to retract beliefs.
  • Interrelations Reveal Structure: The detailed analysis of interrelations (inclusions) between these classes is crucial. It reveals how different definitions of retractability and contraction mechanisms lead to hierarchical relationships, providing a structured understanding of the belief change landscape.
  • Axiomatic Characterization is Key: The use of axiomatic characterizations is a powerful tool. It provides precise, formal definitions for each class of contraction, ensuring clarity and enabling rigorous comparison and analysis.

Key Results

The paper presents axiomatic characterizations for several classes of shielded base contractions, including those induced by:

  • Partial meet contractions.
  • Kernel contractions (two classes).
  • Basic AGM-generated base contractions.

For each of these, they identify five subclasses based on different properties of the "retractable sentences" set. The paper then provides detailed maps illustrating the strict inclusion relationships between these classes, showing how certain combinations of properties lead to more restrictive or more general contraction operators. For instance, Figure 1 shows how different sets of properties for retractable sentences lead to subclasses of a given contraction type (e.g., SPMC, SP-SPMC, CC-SPMC, etc.), while Figure 2 illustrates how different base contraction types (partial meet, kernel, etc.) relate to each other when shielded.

Limitations and Future Work

The paper is an extended abstract, meaning it summarizes a more detailed work. The authors acknowledge that the presented results allow for predicting the behavior of constructed functions and, conversely, for identifying specific constructions for desired properties. Future work could involve exploring the computational complexity of these operators or applying them to specific AI applications where flexible belief revision is critical. The authors also mention that the work is based on propositional logic, and extending these concepts to more expressive logics could be a valuable direction.

Summary updated

BibTeX
@inproceedings{garapa2019shielded,
  author = {Marco Garapa and Eduardo Fermé and Maurício D. Luís Reis},
  title = {{Shielded Base Contraction (Extended Abstract)}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2019},
  doi = {10.24963/ijcai.2019/889},
}
Figure