Learning in Broker Agent

AAAI 1999

TL;DR

This paper tackles the challenge of inaccurate agent capability information in brokered multi-agent systems. They propose a learning-based broker agent that builds "consumer reports" on other agents by observing interactions and using domain ontologies to refine its knowledge, leading to more effective agent cooperation.

What Problem Does This Paper Address?

In multi-agent systems, a common pattern is to use a "broker agent" to facilitate interactions. Agents advertise their capabilities to the broker, and other agents can query the broker for recommendations on who can perform a specific task. The core problem is that the advertised capabilities might not be accurate. Agents, developed independently, might overstate their abilities or lack crucial details. This inaccuracy, even with sophisticated matching mechanisms, hinders effective cooperation. Imagine two car dealerships advertising "all models, lowest price guaranteed" – in reality, one might have a much better selection of colors or superior service. This paper aims to equip broker agents with the ability to learn the true capabilities of other agents, moving beyond simple advertised descriptions.

Methods

The authors propose a broker agent that learns through various channels:

  • Advertisements: Initial information from agents.
  • Interrogation/Testing: The broker actively probes agents about their capabilities.
  • Feedback: Incorporating feedback on past recommendations, both volunteered and requested.
  • "Past Experience": Learning from the broker's own history of interactions and requests.

A key insight is the "law of locality": aspects of an agent's capability that are referenced more frequently will be learned more accurately. The learning process is grounded in a domain ontology, which is structured as a tree-like class hierarchy. The broker uses this ontology to build "service information trees," associating agent capabilities with specific nodes. The broker continuously refines its knowledge by observing agent interactions.

The paper also touches upon exploring collaboration among multiple brokers to avoid single points of failure and bottlenecks.

Datasets

Not discussed in the paper.

Key Insights and Ideas

  • Beyond Advertisements: The core idea is that relying solely on advertised capabilities is insufficient; brokers need to learn the ground truth.
  • "Agent Consumer Reports": The broker acts like a consumer watchdog, building a more reliable profile of each agent's true abilities.
  • Law of Locality: Frequent interaction with an agent's capability leads to more accurate understanding of that capability and its surrounding information.
  • Ontology as a Foundation: Using a domain ontology provides a structured way to organize and refine learned capability information.
  • Collaborative Brokering: Distributing knowledge and learning across multiple brokers can enhance robustness and scalability.

Key Results

Not discussed in the paper. The paper focuses on the proposed approach and its conceptual underpinnings rather than presenting empirical results.

Limitations and Future Work

The authors acknowledge that the broker agent has "limited domain knowledge" initially, relying on a provided domain ontology. Future work would likely involve empirical evaluation of the learning mechanisms, exploring more sophisticated ontology-based learning, and detailing the collaborative brokering mechanisms. The paper also points to the need for robust mechanisms to handle conflicting information and agent deception.

Summary updated

BibTeX
@inproceedings{1999learning,
  title = {{Learning in Broker Agent}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1999},
}
Figure