Impossibility in Belief Merging (Extended Abstract)

Amílcar Mata Díaz, Ramón Pino PérezIJCAI 2018 · DOI

TL;DR

This paper extends the theory of belief merging by introducing a more general framework for representing agents' knowledge and group consensus. It then uses this framework to prove a generalized version of Arrow's impossibility theorem, demonstrating that certain desirable properties for combining beliefs are inherently contradictory.

What Problem Does This Paper Address?

Belief merging is concerned with how to rationally combine information from multiple sources, especially when those sources conflict. This is crucial in many applications like decision-making, medical diagnosis, and data integration. Existing frameworks often represent individual beliefs as propositional formulas and group beliefs as multisets. However, this can be too simplistic. The authors argue that more complex representations of "epistemic states" (individual beliefs) and "profiles" (group beliefs) are needed to capture richer scenarios. They also observe a strong analogy between belief merging and social choice theory, particularly Arrow's impossibility theorem, which states that no voting system can satisfy all desirable properties simultaneously. This paper aims to explore this analogy further by developing an "epistemic" version of Arrow's theorem within a more general belief merging framework.

Methods

The paper introduces a novel, more general framework for belief merging:

  • Epistemic States: Instead of just propositional formulas, individual beliefs are represented as "epistemic states." These are abstract entities that have an associated propositional formula, allowing for richer representations than simple formulas.
  • Vectorial/Functional Profiles: Group beliefs are no longer treated as multisets but as "vectorial" or "functional" profiles. This means the order and identity of agents matter, and the group's information is represented as a function mapping agents to their epistemic states.
  • Epistemic Social Choice Postulates: The authors adapt classical social choice postulates (like the Pareto condition and independence of irrelevant alternatives) into an "epistemic" setting. These postulates are formulated using logical and set-theoretic terms related to epistemic states and merging operators.
  • Representation Theorems: They establish theorems that connect the abstract merging operators to semantic properties, particularly how the group's merged belief relates to individual beliefs and a notion of "group preference" over possible worlds.

Datasets

Not discussed in the paper.

Key Insights and Ideas

  • Richer Representations are Necessary: The paper strongly suggests that simple propositional formulas are insufficient for representing individual beliefs when dealing with complex merging scenarios. Abstract epistemic states and functional profiles are more appropriate.
  • Deep Analogy with Social Choice Theory: The work solidifies the connection between belief merging and social choice theory by showing that impossibility results analogous to Arrow's theorem can be derived in the belief merging context.
  • Generalized Impossibility: The core contribution is a generalized epistemic version of Arrow's impossibility theorem. This theorem demonstrates that even with more expressive representations, certain desirable properties for combining beliefs (like fairness and independence) are fundamentally incompatible.
  • Dictatorship is Unavoidable: The impossibility result implies that any "rational" belief merging operator that satisfies certain fairness and independence conditions must, in essence, be dictatorial. This means the merged belief will be dominated by a single agent's perspective, similar to how a dictator's vote can override others in social choice.
  • Importance of Maximality Condition: The paper highlights that the "maximality condition" (related to how agents rank possibilities) is crucial for the impossibility result to hold. Without it, or with simpler representations, the impossibility might not arise.

Key Results

The paper's main result is Theorem 3 (Main impossibility theorem), which states that any ES basic merging operator satisfying the Standard Domain, Pareto, and Independence conditions is necessarily a dictatorial operator. This is a significant theoretical finding, showing inherent limitations in achieving fair and rational belief aggregation. The paper also provides representation theorems (Theorem 1) that characterize ES merging operators based on abstract assignments, linking syntactic postulates to semantic properties.

Limitations and Future Work

  • Syntactic vs. Semantic Characterization: While the paper provides semantic characterizations for some postulates, a full syntactic characterization of dictatorial operators is suggested as future work.
  • Complexity of Epistemic States: The paper emphasizes the need for complex epistemic states, but the practical implementation and computational cost of such representations are not deeply explored.
  • Specific Representations: The impossibility result holds for various representations of epistemic states (conditionals, ordinal conditional functions, total preorders), but it's noted that simple propositional formulas might not work if the maximality condition is to be preserved. This suggests further investigation into the interplay between representation choice and the impossibility result.

Summary updated

BibTeX
@inproceedings{diaz2018impossibility,
  author = {Amílcar Mata Díaz and Ramón Pino Pérez},
  title = {{Impossibility in Belief Merging (Extended Abstract)}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2018},
  doi = {10.24963/ijcai.2018/799},
}
Figure