Modelling Iterative Judgment Aggregation

Zoi Terzopoulou, Ulle EndrissAAAI 2018 · DOI

Abstract

Judgment Aggregation (JA) is a framework for collective decision-making on logically interconnected issues. While traditionally viewed as a one-shot process, real-world deliberation is often iterative. This paper provides the first formal model for Iterative Judgment Aggregation (IJA), exploring how strategic, myopic agents update their positions. We find that the premise-based rule is exceptionally robust, always converging to socially optimal outcomes, whereas the plurality rule is prone to instability and cycles.

Introduction

In many group settings, members must reach a consensus on a set of related propositions. For instance, a committee might need to decide if a project is on track and if the budget is sufficient to conclude whether the project should continue. Judgment Aggregation (JA) provides the formal tools to handle these dependencies. However, the standard JA model assumes a single round of voting. In practice, agents often observe others' opinions and adjust their own to influence the final result. This paper introduces a mathematical model to study the dynamics and outcomes of such iterative processes.

The Iterative Model

We model a group of agents who are both myopic (focused on the immediate next round) and strategic (willing to change their reported judgments to achieve a more preferred collective outcome). The study focuses on two primary aggregation rules:

  1. Premise-based rule ($F^{pr}$): The group votes on individual premises, and the conclusion is derived logically from those results.
  2. Plurality rule: The group selects the most frequently submitted complete judgment set.

The model also considers different information environments, ranging from full transparency to "full ignorance," where risk-averse agents only move if they are certain of an improvement.

Analysis of the Premise-Based Rule

The premise-based rule demonstrates remarkable stability in an iterative context. It aggregates votes on individual premises to ensure a logically consistent collective conclusion, even when individual agents' judgments might lead to a paradox.

The premise-based rule ($F^{pr}$) applied to three agents, illustrating how individual judgments on premises determine the collective conclusion.

Our analysis proves that under $F^{pr}$, the group is guaranteed to converge to an equilibrium in at most $O(n)$ rounds. Crucially, iteration acts as a corrective mechanism: even if the initial honest profile is socially suboptimal, strategic updates will drive the group toward the socially optimal choice. This results in a Dynamic Price of Anarchy (DPoA) of 1, indicating no loss in social welfare at equilibrium.

The Plurality Rule and Convergence

In contrast, the plurality rule—while simpler to implement—is far less stable. Because it ignores the logical structure of the issues, it is susceptible to infinite cycles where agents continuously shift their votes without reaching a stable state.

A sequence of five rounds under the plurality rule demonstrating a cycle, where the state in round 5 returns to that of round 1.

Furthermore, iteration under the plurality rule does not guarantee an improvement in social welfare. In some scenarios, the group may settle on an equilibrium that is worse than the initial honest outcome, or fail to reach an equilibrium entirely. The "Price of Truth" (PoT) can be high, as agents are incentivized to deviate from their true beliefs without a guaranteed collective benefit.

Information and Strategic Behavior

The dynamics of iteration are heavily influenced by the information available to agents. In Full Information settings, agents see the entire profile of their peers' judgments. In Partial Information settings, they might only see the current winning judgment set. Interestingly, the premise-based rule remains robust even under "full ignorance," where agents are risk-averse and only change their vote if they are certain it won't lead to a worse outcome. This suggests that the logical constraints of the rule guide agents toward the social optimum regardless of the transparency of the process.

Conclusion

This research highlights a fundamental trade-off in iterative decision-making. Rules that respect the logical structure of a problem, like the premise-based rule, benefit from iteration by converging quickly to optimal states. Simpler, popularity-based rules like plurality can become volatile and unpredictable. Future research should explore non-myopic agents who plan multiple steps ahead and the behavior of other complex JA rules in iterative environments.

Summary updated

BibTeX
@inproceedings{terzopoulou2018modelling,
  author = {Zoi Terzopoulou and Ulle Endriss},
  title = {{Modelling Iterative Judgment Aggregation}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2018},
  doi = {10.1609/aaai.v32i1.11440},
}
Figure