Agent Amplified Communication

Henry A. Kautz, Bart Selman, Al MilewskiAAAI 1996 · OpenReview

TL;DR

This paper introduces "agent amplified communication," a framework using intelligent agents to automate and improve person-to-person information gathering, specifically for finding experts within an organization. By analyzing email patterns, these agents can intelligently "referral chain" requests, making it easier to find the right person without overwhelming colleagues.

What Problem Does This Paper Address?

Finding specific information often boils down to two approaches: asking a program or asking a person. While programs (like search engines) are improving, a vast amount of valuable information resides only in people's heads, often due to proprietary or social reasons. The "ask a person" approach, primarily through email or newsgroups, faces challenges: broadcasting to too many people annoys recipients, while targeted emails might miss the right person. Existing systems struggle to bridge the gap between automated information retrieval and the nuanced, often implicit, social networks that hold crucial knowledge. This paper tackles the problem of efficiently and effectively leveraging these informal human networks for information discovery.

Methods

The core of the proposed solution is an agent-based framework. Each user has a "user-agent" that models their areas of interest by analyzing their email communications. These user-agents also build a coarser model of the user's contacts based on their email exchanges. When a user needs information (e.g., to find an expert), they present their query to their user-agent. The agent then intelligently broadcasts this query to the user-agents of potentially relevant contacts, rather than directly to the people themselves. This "agent-to-agent" communication allows for automated filtering and referral. The system uses information retrieval techniques, specifically inverted indexes, to build these models and match queries. The referral process is designed to mimic human "referral chaining," where a person who doesn't know the answer refers you to someone they know who might.

Datasets

Not discussed in the paper. The authors mention using email records collected over a substantial period of time for building user models and contact networks. They also refer to simulations run on randomly generated graphs of varying sizes (100,000 nodes) and connectivity (average neighbors per node).

Key Insights and Ideas

  • Leveraging Implicit Social Networks: The paper highlights that valuable information is often embedded in informal social networks, and agents can be trained to navigate these networks by analyzing communication patterns.
  • Agent-to-Agent Communication: Shifting the burden of initial filtering and referral from humans to their agents significantly reduces the annoyance factor associated with broad email blasts.
  • Trade-off between Responsiveness and Accuracy: Simulations demonstrate a clear trade-off: higher responsiveness (agents are more likely to respond to requests) can compensate for lower referral accuracy (agents might not always point to the best expert), and vice-versa.
  • Privacy is Paramount: Preliminary user experience shows that privacy concerns are critical. Users need to have control over what information their agents share and how they interact.
  • Modeling Human Referral Chains: The agent system is designed to mimic the successful, albeit often inefficient, human process of asking colleagues for referrals, aiming to automate and scale it.

Key Results

  • Simulation Effectiveness: Simulations on a 100,000-node network showed that agent-amplified communication can achieve high success rates (e.g., >50% with R=0.9, A=0.3) with a manageable number of messages (around 200), significantly outperforming a purely random search.
  • Trade-off Demonstrated: Figures and tables illustrate that a high success rate can be achieved even with reduced accuracy (A=0.4) by increasing responsiveness (R=0.9), and conversely, a high responsiveness (R=0.4) cannot compensate for very low accuracy.
  • Preliminary User Experience: A prototype system demonstrated that useful automatic referrals could be found in practice, and that users are willing to engage with such systems, provided privacy concerns are addressed.

Limitations and Future Work

  • Accuracy of Automatic Referrals: The paper acknowledges that automatically generated referrals are generally less accurate than those made by humans.
  • Privacy Concerns: The successful deployment of personal agents hinges on addressing privacy issues, requiring agents to reason about trust and authority.
  • Simulation vs. Real-World Graphs: Current simulations use random graphs. Future work could involve running simulations on graphs derived from actual email communication patterns to better reflect real-world social structures.
  • More Complex Scenarios: For more complex information-gathering tasks, agents will need to be more sophisticated in reasoning about trust and authority.

Summary updated

Abstract

We propose an agent-based framework for assisting and simplifying person-to-person communication for information gathering tasks. As an example, we focus on locating experts for any specified topic. In our approach, the informal person-to-person networks that exist within an organization are used to "referral chain" requests for expertise. User-agents help automate this process. The agents generate referrals by analyzing records of email communication patterns. Simulation results show that the higher responsiveness of an agent-based system can be effectively traded for the higher accuracy of a completely manual approach. Furthermore, preliminary experience with a group of users on a prototype system has shown that useful automatic referrals can be found in practice. Our experience with actual users has also shown that privacy concerns are central to the successful deployment of personal agents: an advanced agent-based system will therefore need to reason about issues involving trust and authority.
BibTeX
@inproceedings{kautz1996agent,
  author = {Henry A. Kautz and Bart Selman and Al Milewski},
  title = {{Agent Amplified Communication}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1996},
  url = {https://openreview.net/forum?id=BJ49UJ-uWB},
}

Loading related content...

Figure