Abstract
Merging information from multiple sources is a fundamental task in artificial intelligence, yet it often encounters the hurdle of conflicting data. This paper addresses the specific challenge of merging prioritized knowledge bases that are individually inconsistent. Traditional methods often suffer from the "drowning problem," where valid low-priority information is discarded due to local inconsistencies. We present two novel approaches—an argument-based method and a multiple-operator method—that preserve useful information and provide a logically sound fusion of contradictory sources.
Introduction
In distributed artificial intelligence and database systems, merging information from multiple agents is essential. Possibilistic logic provides a robust framework for this by attaching a "necessity degree" (a priority weight between 0 and 1) to each logical formula. However, most existing merging methods assume that each individual knowledge base (KB) is self-consistent. When a KB is individually inconsistent, traditional $\pi$-consequence relations calculate an "inconsistency degree" and ignore any formula with a priority lower than that degree. This leads to the "drowning problem," where perfectly valid, non-conflicting information is lost because it was assigned a low weight in a noisy database.
Proposed Methods
1. Argument-Based Merging Method
The authors propose upgrading the traditional $\pi$-consequence relation to an argument-based consequence relation. Instead of a blanket deletion of formulas below the inconsistency threshold, this method retains a formula if a consistent subbase within the KB provides a valid logical "argument" for it. This approach is a strict generalization: it behaves identically to traditional methods when KBs are consistent but successfully rescues low-priority data in inconsistent environments.
2. Multiple-Operator Based Merging Method
A significant flaw in traditional merging is the use of a single mathematical operator for all formulas. Applying a "reinforcement" operator (which increases weight when sources agree) to conflicting statements can artificially inflate certainty. To resolve this, the authors propose a dual-operator approach:
- Consistent Formulas: A strongly conjunctive operator (e.g., probabilistic sum) is used to reinforce belief when multiple sources agree.
- Conflicting Formulas: An up-averaging operator (e.g., weighted average) is used to combine weights without inflating certainty.
A refined version of this method introduces "weakly supported" conflicts to handle cases where a formula is technically in conflict but has significantly more support from one source than others.
Key Insights and Results
This research shifts the perspective on inconsistency, treating it as a natural state of distributed information rather than an error to be purged. By moving from threshold-based cutoffs to argument-based evaluations, the "drowning problem" is effectively bypassed.
Theoretical results and worked examples demonstrate that these methods yield more intuitive outcomes. For instance, agreed-upon formulas see their necessity degrees reinforced, while heavily conflicting formulas see their degrees degrade appropriately—a nuance that single-operator frameworks cannot achieve. Proposition 5 in the paper formally proves that the argument-based method retains formulas that previous state-of-the-art methods would erroneously discard.
Conclusion and Future Work
While the argument-based method excels at preserving information, the multiple-operator method provides more intuitive belief degrees for conflicting data. However, the refined multiple-operator approach is computationally intensive, requiring the calculation of minimally inconsistent subbases and logical closures. Future research will focus on optimizing these logical checks to ensure the framework can scale to massive, real-world knowledge bases.
Summary updated