Case-based Learning from Proactive Communication.

Santiago Ontañón, Enric PlazaIJCAI 2007 · OpenReview

Introduction

In distributed Case-Based Reasoning (CBR) systems, agents solve new problems by referencing similar historical cases. While centralizing data often seems like the most straightforward path to improving performance, it is frequently impractical due to privacy concerns, bandwidth constraints, or the "utility problem"—where excessive redundant data can actually degrade system performance. This paper introduces Proactive Communication Case-based Learning (PCCL), a protocol that enables distributed agents to autonomously improve their performance by bartering specific, high-value cases.

Methodology: Proactive Communication

The PCCL framework moves beyond simple data sharing. Instead, agents identify their "informational deficits" and acquire only the most useful cases from their peers. This is achieved through two primary mechanisms:

1. Justified Predictions

When an agent makes a prediction, it provides a symbolic justification—a description of the features that led to its decision. For example, in identifying a marine sponge, an agent doesn't just name the species; it explains that the classification is based on specific features like the absence of gemmules and the presence of a smooth tylostyle.

Figure 1: An example of a symbolic justification for identifying a sponge species.

2. The Bartering Protocol

Agents engage in a proactive 1:1 barter system. When an agent receives a new case, it asks peers for predictions. If a peer provides an incorrect answer, the first agent identifies counterexamples in its own database—cases that match the peer's incorrect justification but have a different, correct solution. These counterexamples form "refutation sets." By exchanging exactly one case per refutation set, agents ensure that they only receive data addressing specific weaknesses, maintaining a compact and efficient case base.

Experimental Evaluation

The researchers evaluated PCCL using a 5-agent system across three distinct datasets: Marine Sponges (280 cases), Soybean plant disease diagnosis (307 cases), and Zoology animal classification (101 cases).

Results and Discussion

The experimental results demonstrate that Learning from Communication (LFC) consistently outperforms Individual Learning (IL) across all tested domains.

Figure 6: Accuracy comparison between Learning from Communication (LFC) and Individual Learning (IL) across three datasets.

Key findings include:

  • Enhanced Accuracy: In the Zoology dataset, LFC reached approximately 82% accuracy with significantly fewer cases than required by IL agents.
  • Efficiency and Redundancy Control: In the Sponges dataset, LFC achieved 82.07% accuracy with only 50.4 retained cases. This outperformed agents that had access to 100% of the training data but lacked the bartering protocol (79.21% accuracy).
  • The "Less is More" Principle: The strict 1:1 bartering policy proved superior to looser $n:m$ policies. By being selective, agents avoided bloating their databases with redundant information that did not contribute to better decision-making.

Conclusion

PCCL demonstrates that justifications can serve as powerful diagnostic tools in multi-agent systems. By proactively testing peers and bartering for counterexamples, agents can achieve higher accuracy with smaller, more efficient case bases. Future work may explore more complex negotiation strategies and address potential issues with non-honest agents in competitive environments.

Summary updated

BibTeX
@inproceedings{ontanon2007case,
  author = {Santiago Ontañón and Enric Plaza},
  title = {{Case-based Learning from Proactive Communication.}},
  booktitle = {IJCAI},
  year = {2007},
  url = {https://openreview.net/forum?id=ks5Kotc_m_J9},
}
Figure