Multiparty Proactive Communication: A Perspective for Evolving Shared Mental Models

Kaivan Kamali, Xiaocong Fan, John YenAAAI 2006

Abstract

Effective teamwork in multi-agent systems requires a Shared Mental Model (SMM)—a mutual understanding of team goals and the current state of the world. Traditional Agent Communication Languages (ACLs) often fail to maintain this synchronization, relying on either restrictive point-to-point dialogues or overly broad broadcasts. This paper introduces a formal framework for "multiparty proactive communication," enabling agents to communicate with specific addressees while treating overhearers as first-class citizens. By formalizing how agents anticipate teammates' needs and share inference logic, the framework allows for proactive assistance and more robust team coordination.

Introduction: The Challenge of Shared Mental Models

In both human and artificial teams, high performance is predicated on a Shared Mental Model. However, existing communication protocols like KQML or FIPA are primarily designed for two-party interactions. Point-to-point communication often leaves the rest of the team uninformed, leading to a fractured SMM. Conversely, broadcast communication can be too vague, lacking a designated addressee and often resulting in the "bystander effect," where no single agent takes responsibility for a request.

The authors address this gap by proposing a formal logic that allows overhearers to update their beliefs about what their teammates know and need. This enables "proactive" behavior, where an agent provides information before it is explicitly requested, based on its understanding of the team's evolving mental state.

Formalizing Multiparty Communication

The framework is built upon SharedPlans theory and Joint Intentions Theory, utilizing modal logic to define the mental states (Beliefs, Desires, and Intentions) of agents. The core of the method involves three novel communication acts, or "performatives":

  1. MP-Inform (Multiparty Inform): An extension of standard inform acts where the speaker, addressee, and all overhearers become mutually aware that a message has been shared.
  2. MP-ProInform (Multiparty Proactive Inform): Triggered when an agent anticipates a teammate's information need and provides the data preemptively.
  3. MP-IndProInform (Multiparty Indirect Proactive Inform): The most sophisticated act, where an agent shares not just a fact, but the inference logic required to reach a conclusion. For example, if an agent knows that Enemy Close + Attack Pattern = High Threat, it can share this rule so that an overhearer with partial data can complete the puzzle for the team.
Figure 1: A scenario demonstrating how Agent C's needs and Agent S1's inference rules are shared across the team via multiparty communication.

Key Insights: Overhearing and Inference

A central insight of this research is that overhearing is not a side effect but a primary mechanism for situational awareness. In human teams, such as flight crews, listening to others' conversations is vital for synchronization. This framework treats overhearers as active participants who use shared information to update their internal models of their teammates.

Furthermore, the paper highlights the power of sharing "inference knowledge." Agents do not just need raw data; they need to understand the rules their teammates use to process that data. By sharing an inference rule (e.g., $b \land d \Rightarrow a$), an agent enables an overhearer who might possess piece $b$ to realize they can now help a teammate who needs $a$.

Results and Scenario Analysis

The authors validate their framework through formal theorems and a "Commander and Scouts" scenario. The theorems prove that these multiparty acts successfully establish mutual belief across the entire group and are sufficient to trigger "helping" intentions in overhearers without explicit commands.

Figure 2: The Commander and Scouts scenario illustrating how agents track mutual ignorance (~Knows) to trigger proactive information sharing.

In the illustrated scenario, Agent C (Commander) expresses a need for information ($a$). Through multiparty communication, Scouts S1 and S2 become aware of this need and each other's knowledge. Even if S1 only has part of the required information, sharing the logic allows S2 to proactively provide the missing piece, ensuring the Commander's need is met efficiently.

Limitations and Future Work

While the theoretical foundation is rigorous, several challenges remain for real-world application:

  • Communication Reliability: The current logic assumes messages are always delivered, which is rarely true in mobile or peer-to-peer environments.
  • Information Overload: If every agent processes every overheard message, the system may suffer from excessive "noise" and computational overhead.
  • Implementation Scale: Future work is needed to test these protocols in large-scale software implementations beyond the initial CAST (Collaborative Agents for Simulating Teamwork) framework simulations.

Ultimately, this work provides a mathematical foundation for proactivity, moving beyond heuristic-based approaches to a rigorous logic of team coordination.

Summary updated

BibTeX
@inproceedings{kamali2006multiparty,
  author = {Kaivan Kamali and Xiaocong Fan and John Yen},
  title = {{Multiparty Proactive Communication: A Perspective for Evolving Shared Mental Models}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2006},
}
Figure