On the Aggregation of Argumentation Frameworks
Introduction
In multi-agent systems, reaching a consensus when individual agents hold conflicting views is a central challenge. This paper explores the aggregation of abstract argumentation frameworks (AFs), a standard model in Artificial Intelligence for representing debates. In an AF, arguments are nodes in a graph, and conflicts between them are represented as directed "attack" edges. While methods exist to determine "winning" arguments within a single framework, aggregating multiple individual frameworks into a single "social" consensus remains a complex problem.
The Challenge of Aggregation
When multiple experts provide their own AFs, we need a way to merge them that respects certain "rationality properties"—mathematical standards for fairness and logical consistency. For instance, if every agent agrees that one argument attacks another, that attack should ideally appear in the collective result. Previous work by researchers like Dunne et al. established a set of such properties, but until now, few studies had rigorously tested existing aggregation methods against them. This paper addresses this gap, revealing a fundamental tension between the structural representation of a debate and its eventual outcome.
Methods and Operators
The authors evaluate three primary categories of aggregation operators:
- Distance-based Merging: These operators, such as $\Delta^\Sigma_{de}$, seek a consensus graph that minimizes the total Hamming distance from the set of individual graphs. They focus on preserving the overall structure of the attacks.
- Voting-based Methods: These rely on qualified voting (QV), where an attack edge is included in the social graph only if it receives support from a specified threshold of agents (e.g., a simple majority).
- Weighted Argumentation Frameworks (WAFs): The authors introduce three new operators—$FUS_{All}$, $FUS_{AllNT}$, and $FUS_{MajNT}$. These treat the number of agents supporting an attack as a weight. They employ specialized algorithms to "relax" the graph or select optimal sets of winning arguments based on these weights, specifically aiming to avoid "trivial" results where no arguments are accepted.
Key Insights: Structure vs. Outcome
A major contribution of this work is identifying the "rationality gap." Many existing operators fail to satisfy basic logical properties. A core issue is the divide between properties focusing on the structure of the graph (the attacks) and those focusing on the extensions (the sets of winning arguments).
As shown in the example above, different frameworks can lead to the same set of winning arguments despite having different internal conflicts. A merging method that is "fair" to the arrows in the graph might produce a result where the set of winning arguments is nonsensical or empty—the "Non-Triviality" problem. The authors' proposed $FUS_{AllNT}$ and $FUS_{MajNT}$ operators are designed to ensure the group always reaches a non-empty conclusion.
Results and Evaluation
The paper provides a comprehensive evaluation of these methods. The results show that:
- Distance-based operators are effective at preserving argument identity but perform poorly on extension-related properties.
- Qualified Voting is generally weak, failing most rationality tests.
- WAF-based operators ($FUS_{MajNT}$) show significant improvement, particularly in ensuring non-triviality and preserving identity when agents are in complete agreement. However, even these struggle with strict "Majority" properties, as the internal logic of argument defense can override simple vote counts.
Conclusion and Future Work
The authors conclude that traditional rationality properties from social choice theory may be too restrictive for argumentation. Unlike political candidates, arguments have a complex defense structure that must be accounted for. Future research will focus on developing new postulates tailored to the unique structure of argumentation and exploring how these operators can be applied to real-world negotiation protocols.
Summary updated