Multi-Agent Abstract Argumentation Frameworks With Incomplete Knowledge of Attacks

Andreas Herzig, Antonio Yuste-GinelIJCAI 2021 · DOI

TL;DR

This paper introduces a novel framework for multi-agent argumentation that models agents' incomplete knowledge of attack relations between arguments. By leveraging dynamic epistemic logic, it allows for reasoning about how agents' beliefs about these attacks evolve during a debate, enabling more realistic modeling of strategic interactions.

What Problem Does This Paper Address?

Traditional abstract argumentation frameworks (AFs) model arguments and their conflicts but often assume agents have complete knowledge of these relationships. This is unrealistic for several reasons: agents might have limited reasoning capabilities and fail to identify all actual attacks, or arguments might be presented implicitly (enthymemes), leaving some attacks unstated. Furthermore, real-world debates involve multiple agents interacting, each with their own potentially incomplete view of the argumentative landscape. Existing work has addressed multi-agent systems and dynamics separately, but this paper aims to unify these aspects by modeling agents' partial knowledge of attack relations specifically, and how this knowledge changes over time through public announcements.

Methods

The core contribution is a multi-agent, dynamic extension of abstract argumentation frameworks (AFs) grounded in epistemic logic. Key methodological aspects include:

  • Focus on Attack Incompleteness: Unlike many approaches that model incomplete knowledge of arguments, this work focuses on agents having partial knowledge of the attack relation itself. An agent's "view" of an AF is represented as a subgraph of the actual attack relation.
  • Second-Order Knowledge: The framework incorporates "opponent models" by allowing agents to have knowledge about other agents' knowledge of attacks (e.g., "agent A knows that agent B knows that argument X attacks argument Y").
  • Public Views: A "public view" of the debate is maintained, representing common knowledge about the attack relation, which evolves as agents make public announcements.
  • Dynamic Epistemic Logic (S5-PAL): The entire framework, including the evolution of agents' views and the public view through announcements, is formally described using S5-PAL, a well-established dynamic-epistemic logic. This allows for precise modeling of knowledge updates and their consequences.
  • Epistemic Characterization: AFs and their extensions are characterized using epistemic logic formulas. For instance, an agent's knowledge of an attack (a, b) is represented by a formula like Kira,b.

Datasets

Not discussed in the paper.

Key Insights and Ideas

  • Attack-centric Incompleteness is Crucial: Modeling incomplete knowledge of attacks, rather than just arguments, is a more natural fit for scenarios like enthymematic arguments or agents with bounded rationality.
  • Strategic Reasoning via Opponent Models: The inclusion of second-order knowledge (knowing what others know) is essential for agents to strategize effectively in a debate, anticipating how their actions might be perceived or countered.
  • Public Announcements Drive Debate Dynamics: The framework elegantly captures how public announcements (e.g., an agent declaring an attack) update not only the public state of the debate but also individual agents' knowledge and views.
  • Epistemic Logic as a Unifying Language: S5-PAL provides a powerful and expressive language to formally define and reason about these complex multi-agent, dynamic, and knowledge-rich argumentation scenarios.

Key Results

The paper provides formal characterizations of multi-agent AFs and their updates using epistemic logic. Key results include:

  • Characterization of Multi-Agent AFs: Formulas are provided to represent an agent's first-order view (Kira,b), second-order view (KjKira,b), and the public view (Cagtra,b) of the attack relation.
  • Characterization of Extensions: A formula (Stablepub) is derived to characterize public stable extensions of the multi-agent AF, allowing for reasoning about public acceptance of arguments.
  • Characterization of Updates: Propositions show how public announcements of attacks by an agent i (e.g., i:(a,b)) translate into updates in the epistemic logic model, affecting agents' views and the public view. For example, (a, b) ∈ Ri∪{(a, b)} iff 0(M) → [Kira,b!]Kjrc,d is S5-PAL valid.

Limitations and Future Work

  • Higher-Order Views: The paper discusses extending the framework to arbitrary higher-order views (e.g., knowing what someone knows about what someone else knows about an attack), suggesting this can be achieved by directly using S5 models.
  • Uncertainty: The current framework models knowledge (certainty) but not belief or uncertainty about attacks. The authors suggest combining their approach with existing work on incomplete argumentation frameworks (IAFs) to model attacks that agents are uncertain about.
  • Dialogue Modeling: A promising future direction is to define explicit dialogue protocols based on these multi-agent AFs and apply the framework to compute optimal next moves in a debate.
  • Belief Revision: Extending the framework to handle belief revision, beyond just knowledge updates, is identified as an open question for modeling argumentative/doxastic dynamics.

Summary updated

BibTeX
@inproceedings{herzig2021multi,
  author = {Andreas Herzig and Antonio Yuste-Ginel},
  title = {{Multi-Agent Abstract Argumentation Frameworks With Incomplete Knowledge of Attacks}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2021},
  doi = {10.24963/ijcai.2021/265},
}
Figure