An Egalitarist Fusion of Incommensurable Ranked Belief Bases under Constraints

Salem Benferhat, Sylvain Lagrue, Julien RossitAAAI 2007

The challenge of merging conflicting information from multiple sources is central to automated reasoning and artificial intelligence. When experts or sensors provide ranked beliefs—indicating not just what they believe, but how strongly they believe it—the fusion process becomes complex. A significant hurdle in this domain is incommensurability: the reality that different sources often use incomparable priority scales. This paper presents a novel "egalitarian" approach to fusion that respects internal rankings while avoiding arbitrary assumptions about how those scales relate to one another.

The Problem of Incommensurable Beliefs

Standard belief fusion techniques often assume commensurability, implying that a "Rank 8" from Expert A carries the same weight as a "Rank 8" from Expert B. In practice, internal scales are subjective. One expert's "highly likely" might correspond to another's "somewhat likely."

The authors address this by focusing on Ranked Belief Bases, where information is structured using Spohn’s Ordinal Conditional Functions (OCF), representing "degrees of surprise." The goal is to merge these bases under hard constraints—truths that must be satisfied—while ensuring the result is "egalitarian." Unlike utilitarian methods that might favor a majority at the expense of a minority, egalitarian fusion seeks to minimize the maximum dissatisfaction among all participating agents.

Egalitarian Fusion and the Max Operator

To achieve fairness, the proposed method utilizes the max operator to represent the collective dissatisfaction for any given interpretation of the world. By identifying the models that minimize this maximum dissatisfaction, the system ensures that no single agent is disproportionately ignored.

Example of egalitarian fusion using the Max operator to determine dissatisfaction across possible worlds.

As shown in the example above, the fusion operator evaluates each possible world ($\omega$) by taking the maximum surprise value across the individual belief bases ($B_1, B_2$). The final fused result selects the worlds where this maximum value is minimized (in this case, $\omega_2$ with a Max value of 2).

Handling Incommensurability via Scalings

The core innovation of the paper lies in how it handles the lack of a common scale. The authors introduce the concept of Compatible Scalings: re-assignments of numerical ranks that preserve the original relative ordering of a source's beliefs.

Illustration of different compatible scalings ($\mathcal{S}_1, \mathcal{S}_2$) for ranked formulas.

By considering all possible compatible scalings, the authors define a "skeptical" merging operator. A solution is preferred only if it remains the optimal choice regardless of which valid numerical scale is applied to the experts' reports. This prevents the fusion result from being an artifact of an arbitrary choice of scale.

The Pareto Equivalence

The most significant theoretical contribution is Proposition 16. The authors prove that the computationally intensive task of checking every possible compatible scaling is mathematically equivalent to a much simpler Pareto-like ordering.

Under this ordering, one interpretation is considered better than another if it is less surprising for at least one expert and not more surprising for any others. This "Aha!" moment transforms a philosophically rigorous requirement—skepticism over all possible scales—into a practical, algorithmic approach for belief fusion.

Key Results and Properties

The proposed operator satisfies several critical "IC postulates" (rationality properties for merging), including IC0 through IC5, IC7, and IC8. Notably, it adheres to Majority Independence (MI). In an egalitarian framework, a large group of experts cannot simply "outvote" a lone dissenter if doing so would cause extreme dissatisfaction for that individual. This prevents a "tyranny of the majority" within the knowledge base.

Furthermore, the framework distinguishes between "beliefs" (which can be weakened to resolve conflict) and "constraints" (hard truths). The fusion result is guaranteed to be a subset of the models that satisfy these hard constraints.

Limitations and Future Outlook

While robust, the method has limitations. It fails the IC6 postulate, a known trait of max-based egalitarian operators. Additionally, the skeptical nature of the operator can sometimes lead to very "cautious" results—if experts disagree wildly, the fused output might be less informative (closer to a tautology).

Future research could delve into the computational complexity of applying this Pareto characterization in large-scale propositional logic spaces and explore more "informative" variations of skeptical fusion that still maintain egalitarian principles.

Summary updated

BibTeX
@inproceedings{benferhat2007egalitarist,
  author = {Salem Benferhat and Sylvain Lagrue and Julien Rossit},
  title = {{An Egalitarist Fusion of Incommensurable Ranked Belief Bases under Constraints}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2007},
}
Figure