Abstract
In loosely-coupled, extended-duration teamwork (ISLET), coordinating information sharing without a master plan or overwhelming team members is a significant challenge. This paper introduces Mutual Influence Potential Networks (MIP-Nets), a framework that implicitly learns collaboration patterns and task dependencies from team interactions. By calculating a Degree of Interest (DOI) based on role proximity and task structure, the MIP-DOI algorithm identifies relevant information to share. Evaluations on Wikipedia revision histories and collaborative simulations demonstrate that MIP-Nets significantly outperform standard baselines, approaching the performance of omniscient systems in focused scenarios.
Introduction: The ISLET Challenge
In complex, distributed teamwork—such as multidisciplinary medical teams treating chronic conditions or researchers co-authoring large-scale grants—coordination is often hindered by the "loosely coupled" nature of the work. Members operate independently on distinct project components over extended periods. This environment, termed ISLET (Information Sharing in Loosely-coupled Extended-duration Teamwork), presents three primary hurdles:
- Absence of a Master Plan: Unlike traditional multi-agent systems, there is no pre-defined plan library or complete state-space model.
- Information Overload: Sharing all updates is cognitively taxing, while sharing too little leads to coordination failures.
- Dynamic Dependencies: Task relationships evolve, making it difficult to discern which updates are relevant to specific team members.
Existing AI models for information sharing typically rely on the "complete plan knowledge assumption." This research moves beyond that constraint, focusing on learning relevance by observing team behavior.
The MIP-Net Framework
The authors propose a two-part solution: Mutual Influence Potential Networks (MIP-Nets) and the MIP-DOI algorithm.
1. Network Representation
MIP-Nets use a graph-based structure where nodes represent both partners (team members) and objects (tasks, documents, or records).
- Role Allocation: Edges between a partner and an object represent the frequency and nature of their interactions, implicitly defining roles.
- Task Structure: Edges between objects represent functional dependencies; if Object A and Object B are frequently modified together, a strong edge forms between them.
2. The MIP-DOI Algorithm
To filter information, the system calculates a Degree of Interest (DOI) for each object relative to a partner. This score is a weighted combination of:
- A Priori Importance (Centrality): The object's overall significance within the project graph.
- Proximity to Partner: The strength of the relationship between the object and the partner's established role.
- Proximity to Focus: The relationship between the update and the partner's current active task.
Experimental Evaluation
The approach was evaluated using two distinct datasets:
- Wikipedia Revision Histories: Analyzing 24 "featured" articles to determine if MIP-Nets could predict future edits based on past author-paragraph interactions.
- Collaborative Activity Simulation: A graph-coloring simulation where agents coordinated under constraints, providing a ground truth for measuring the relevance of shared information.
Performance Analysis
The results indicate that MIP-Nets effectively bridge the gap between simple heuristics and omniscient knowledge.
- Wikipedia Analysis: MIP-DOI configurations consistently outperformed baselines such as "Most Recently Changed" or "Most Frequently Changed." The "centrality" variant was particularly effective in this decentralized setting, achieving a 26% precision improvement over random baselines.
- Simulation Success: In the graph-coloring task, the MIP-DOI-focus variant—which incorporates the agent's current task—performed nearly as well as an "Omniscient" baseline that possessed full knowledge of task constraints.
- Efficiency: The system remained effective even under strict communication budgets (sharing only 1-3 items), successfully filtering noise without losing critical coordination data.
Discussion and Future Directions
MIP-Nets demonstrate that roles and task structures can be learned implicitly, solving the "cold-start" problem of coordination in new teams. However, the system currently assumes relatively stable collaboration patterns. Future work aims to:
- Incorporate Feedback: Allowing users to explicitly flag irrelevant information to accelerate network refinement.
- Weight Decay: Implementing "forgetting" mechanisms to handle evolving project structures and role changes.
- Hybrid Initialization: Using partial domain knowledge (e.g., medical ontologies) to improve performance in the early stages of a project.
By treating task structure as a proxy for formal plans, MIP-Nets provide a scalable, low-overhead solution for information sharing in complex, long-term human-AI teams.
Summary updated