Distributive and Collective Readings in Group Protocols

Silvia Rossi, Sanjeev Kumar, Philip R. CohenIJCAI 2005

Introduction

In multi-agent systems (MAS), coordination is the cornerstone of successful task execution. Whether managing autonomous swarms or grid computing resources, agents must communicate to align their goals. However, most traditional communication protocols are fundamentally bilateral, designed for one-on-one interactions. When these protocols are scaled to groups, they often treat group communication as a mere shortcut—sending the same message to multiple individuals simultaneously.

This paper argues that such an approach overlooks a critical semantic distinction: the difference between distributive and collective readings. A distributive request (e.g., "everyone lift a chair") results in isolated, pair-wise commitments. In contrast, a collective request (e.g., "lift this heavy table together") creates a unified team commitment. Without a formal logic to distinguish these, agents cannot reliably form cohesive teams characterized by mutual belief and joint persistence.

Theoretical Framework

The researchers develop their framework using Joint Intention Theory, employing modal logic to define the mental states of agents, including their Beliefs, Goals, and Commitments.

Key technical components include:

  • Persistent Weak Achievement Goal (PWAG): A formalization of a directed commitment from one agent to another.
  • Mutual Belief (MB): A state where agents not only share a belief but also believe that all other members of the group share that belief.
  • The "Underline" Notation: The authors introduce a syntactic distinction where an underlined group name ($\underline{\gamma}$) denotes a distributive reading (acting as individuals), while a standard group name ($\gamma$) denotes a collective reading (acting as a unit).

Distinguishing Readings

The fundamental difference between a collective and distributive reading lies in the establishment of mutual belief among the recipients. In a collective request, every agent is aware that every other agent has received the same request, forming the basis for a potential team.

Comparison of distributive (a) and collective (b) communication structures.

As illustrated in the figure above, distributive communication (a) establishes independent links between the requester and each participant. Collective communication (b) treats the group as a single entity (indicated by the dashed boundary), where the interaction is between the requester and the group as a whole.

Protocol Semantics and Results

A major contribution of the paper is Theorem 3.3, which proves that sending $N$ individual messages to $N$ agents is not semantically equivalent to a single group message. In the individual case, Agent A may not know that Agent B also received the request, preventing the formation of a Joint Persistent Goal (JPG)—the "glue" that holds a team together.

To handle the complexities of real-world interaction, the authors define the "agreeing group" ($\gamma_{agree}$) retrospectively. This allows for dynamic team formation where the final group consists only of those who responded with an AGREE before a specific deadline.

Mapping of request and agreement types to distributive and collective readings.

The table above summarizes how different combinations of group communicative acts lead to specific readings. For instance, a collective group request paired with a collective agreement (row c) ensures a collective reading, whereas distributive requests (row a) result in independent pair-wise commitments, which are more suitable for competitive scenarios like independent contract bidding.

Conclusion and Future Work

By formalizing these readings, the paper provides a more robust foundation for group protocols. Collective readings ensure that if an agent in a team fails, the mutual belief structure triggers recovery behaviors, such as informing others of the failure, making the team more resilient than a mere collection of individuals.

Future research will address the impact of "over-hearers" (similar to the CC field in emails), the logic required for authorized representatives to speak on behalf of a group, and formal protocols for the graceful disbanding of teams once a task is completed.

Summary updated

BibTeX
@inproceedings{rossi2005distributive,
  author = {Silvia Rossi and Sanjeev Kumar and Philip R. Cohen},
  title = {{Distributive and Collective Readings in Group Protocols}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2005},
}
Figure