On Paraconsistent Belief Revision in LP

Nicolas Schwind, Sébastien Konieczny, Ramón Pino PérezAAAI 2022 · DOI

Introduction

In classical logic, the "principle of explosion" (ex falso quodlibet) dictates that from a contradiction, any statement can be derived ($P \land \neg P \vdash Q$). This poses a significant challenge for artificial intelligence agents operating in real-world environments, where data sources—such as merged web ontologies or conflicting sensor inputs—are frequently inconsistent. Standard belief revision frameworks, most notably the AGM (Alchourrón, Gärdenfors, and Makinson) postulates, are built upon classical foundations and thus collapse when faced with such contradictions. This paper addresses this limitation by adapting belief revision principles to Priest’s Logic of Paradox (LP), a paraconsistent logic designed to handle inconsistencies gracefully.

The Challenge of Revision in LP

Priest’s LP is a three-valued logic that introduces a third truth value, "Both" (B), alongside True (1) and False (0). While LP prevents logical explosion, it introduces a new problem for belief revision: every formula is technically "consistent" because one can always assign the value 'B' to all variables. In a standard AGM setting, this trivializes the revision process, turning every update into a simple conjunction and failing to provide a rational mechanism for choosing between competing beliefs.

Methodology: Redefining Belief Change for LP

The authors bridge the gap between paraconsistency and rational belief change by redefining the fundamental building blocks of the framework:

1. Strong LP Inference and Representative Models

Rather than relying on standard LP consequence, the authors utilize "stronger" versions of inference that focus on representative models. These are the interpretations that are "most classical" while still satisfying a given formula, effectively minimizing the use of the 'B' value unless necessary.

2. The LP+expansion Operator

In classical logic, expansion (adding information that does not conflict with existing beliefs) is equivalent to a logical conjunction. The authors demonstrate that this fails in LP and propose a unique "LP+expansion" operator. This operator selects specific representative worlds to avoid the triviality inherent in 3-valued semantics.

Example 3: Representation of possible worlds in the 3-valued LP logic, where 'B' denotes a statement that is both true and false.

3. Adapted Postulates (LP1-LP6)

The classic KM (Katsuno and Mendelzon) postulates for finite belief revision are adapted into a new set of rules, LP1-LP6. These postulates govern how an agent should rationally update its belief state when new, potentially conflicting information arrives.

Key Results

Representation Theorem

The authors prove a mathematical equivalence showing that any revision operator satisfying the LP1-LP6 postulates can be characterized by a plausibility preorder over possible worlds. This allows agents to rank states of the world by likelihood, providing a semantic basis for revision.

Distance-Based Operators

The paper generalizes Dalal’s revision—a method using Hamming distance—to the 3-valued setting. By introducing a tunable parameter ($d_{B \leftarrow 01}$), the authors allow for the creation of different agent "personalities." Some agents may be "consistency-seeking," striving for classical interpretations at all costs, while others are "contradiction-tolerant," preferring to accept a local inconsistency (assigning 'B') rather than discarding a long-held belief.

Discussion and Future Work

One of the most significant insights of this work is the "isolation of inconsistency." Unlike classical revision, which might force an agent to drop a belief entirely to maintain consistency, this LP-based approach allows the agent to mark a specific variable as inconsistent. This prevents a conflict regarding one topic from "poisoning" unrelated beliefs.

While this framework provides a robust foundation for paraconsistent revision in LP, future work is needed to explore other paraconsistent logics (such as the 4-valued Belnap-Dunn logic) and to address the computational complexity of these operators in large-scale knowledge bases. Additionally, the authors suggest extending this framework to belief merging, where multiple agents' conflicting views must be integrated into a single coherent state.

Summary updated

BibTeX
@inproceedings{schwind2022paraconsistent,
  author = {Nicolas Schwind and Sébastien Konieczny and Ramón Pino Pérez},
  title = {{On Paraconsistent Belief Revision in LP}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2022},
  doi = {10.1609/aaai.v36i5.20532},
}
Figure