TL;DR
This paper provides a formal characterization of a specific belief revision axiom (Parikh's axiom (P)) using the concept of epistemic entrenchment. This work bridges the gap between abstract axiom systems and concrete models of belief change, offering a deeper understanding of how rational agents revise their beliefs locally when faced with new information.
What Problem Does This Paper Address?
The field of belief revision, particularly the AGM (Alchourrón, Gärdenfors, Makinson) paradigm, provides a formal framework for how an agent should update their beliefs when presented with new information. However, the standard AGM postulates are criticized for being too liberal regarding "relevance" – they can lead to agents revising beliefs that are seemingly unrelated to the new information. Parikh's axiom (P) was introduced to address this by enforcing a "local" belief revision, where only beliefs relevant to the new information are affected. While this axiom has been studied, its characterization in terms of epistemic entrenchments (a way to rank beliefs by their importance or "entrenchment") for general, potentially incomplete, belief sets was missing. This paper fills that gap, providing the epistemic-entrenchment conditions that precisely capture the weak version of Parikh's axiom (wP).
Methods
The paper leverages existing formalisms in belief revision:
- AGM Paradigm: The foundational framework for belief revision, defined by a set of rationality postulates (K1-K8).
- Epistemic Entrenchment: A binary relation
<on sentences, representing their relative importance. A set of postulates (EE1-EE5) defines a valid epistemic entrenchment preorder. A key connection is the (<*) condition, which links epistemic entrenchments to AGM revision functions. - Faithful Preorders: A model based on preorders over possible worlds, which also induce AGM revision functions.
- K-Splitting: A novel concept introduced by Parikh, which partitions the propositional variables of a language based on a theory K. This allows for a more granular analysis of relevance. The "finest K-splitting" is crucial for the paper's analysis.
- Support Sets and Cuts: These are defined based on the units of K (derived from the K-splitting) and are used to formalize the notion of "justification" for beliefs. A "cut" represents a minimal set of units whose removal would leave a sentence unsupported.
The core of the paper is to translate the possible-world characterization of (wP) into equivalent conditions on epistemic entrenchments. This involves defining notions of "better supported" and "equally supported" sentences based on support sets and cuts, and then establishing a formal link between these and the epistemic entrenchment preorders.
Datasets
Not discussed in the paper.
Key Insights and Ideas
- Formalizing Local Change: The paper provides the first epistemic-entrenchment characterization of Parikh's weak axiom (wP) for general incomplete theories. This is a significant step towards understanding how rational agents can perform belief revision that respects relevance.
- K-Splitting as a Foundation for Relevance: The concept of K-splitting is central to defining relevance in this context. It allows for a principled way to decompose a theory and analyze how new information affects specific "compartments" of belief.
- Bridging Worlds and Entrenchments: The paper successfully connects the abstract notion of epistemic entrenchment with the more concrete possible-world semantics for belief revision, specifically for the axiom (wP). This demonstrates the power of epistemic entrenchments as a unifying concept.
- Intuitive Conditions for (wP): The derived conditions (EP1) and (EP2) on epistemic entrenchments are presented as natural constraints on the relative "retractability" of sentences, aligning with the intuition that more entrenched beliefs are harder to give up.
- Connection to Ensconcement: The authors highlight that their epistemic-entrenchment characterization aligns well with the concept of "ensconcement" (a preorder on belief bases), suggesting practical implications for implementing belief revision systems.
Key Results
The paper's main result is Theorem 1, which establishes an equivalence between two sets of conditions:
- Possible-World Conditions (Q1, Q2): These conditions, based on an extended definition of "difference" (Diff) that utilizes K-splitting, characterize (wP) for general incomplete theories.
- Epistemic Entrenchment Conditions (EP1, EP2): These conditions, defined using notions of "better supported" and "equally supported" sentences derived from support sets and cuts, are shown to be equivalent to (Q1) and (Q2) when linked via the standard (SE) condition.
Essentially, (EP1) and (EP2) are the epistemic-entrenchment counterparts of (wP) for incomplete theories.
Limitations and Future Work
- Focus on Weak (wP): The paper focuses on the weak version of Parikh's axiom (wP). A natural extension would be to characterize the strong version (sP) in terms of epistemic entrenchments.
- Theoretical Focus: While the paper discusses practical implications for implementing belief revision systems, the work is primarily theoretical, providing formal characterizations.
- Generalization: The authors suggest that the concepts and results might be applicable to other AI domains where relevance and local change are important, such as description logics, ontology updates, and multi-agent systems.
Summary updated