The Multi-Purpose Incremental Learning System AQ15 and Its Testing Application to Three Medical Domains

Ryszard S. Michalski, Igor Mozetic, Jiarong Hong, Nada LavracAAAI 1986

Abstract

AQ15 is an inductive learning system designed for incremental learning and rule induction from noisy data. This paper evaluates AQ15 on three medical diagnostic domains, demonstrating that its unique TRUNC method—which simplifies rule sets by removing rare-case complexes—can maintain or even improve diagnostic accuracy while significantly reducing complexity. Results show that AQ15 matches or exceeds human expert performance in specific medical contexts, proving the efficacy of logic-based machine learning for automated knowledge acquisition.

Introduction

In the mid-1980s, the "knowledge acquisition bottleneck" was a primary obstacle to advancing Artificial Intelligence. While human experts are proficient at decision-making, they often struggle to articulate the underlying logic in a structured format suitable for computer systems. AQ15 was developed to address this by automating rule creation from examples, supporting incremental learning, and handling the inherent noise and complexity of real-world data, particularly in medical diagnostics.

Methods

AQ15 utilizes a heuristic search to induce logical expressions that cover positive examples while excluding negative ones. It employs Variable-Valued Logic ($VL_1$), a flexible propositional calculus allowing variables to take multiple values, and constructive induction to generate new attributes.

A central innovation is the TRUNC method. This technique weights rules (complexes) based on their coverage. "Light" complexes—those explaining only a few rare instances—are truncated to simplify the knowledge base. When strict matching fails due to truncation, AQ15 uses a flexible "measure of fit" to find the most similar concept, effectively treating truncated rules as prototypes.

Figure 1: Illustration of complexes (Cpx) and their coverage levels, which informs the TRUNC simplification process.

Experimental Setup

The system was evaluated on three medical datasets from the Institute of Oncology in Ljubljana:

  1. Lymphography: 148 patients, 4 classes, 18 attributes.
  2. Breast Cancer Recurrence: 286 patients, 2 classes, 9 attributes.
  3. Location of Primary Tumor: 339 patients, 22 classes, 17 attributes.

These domains range from clear patterns (Lymphography) to highly sparse and overlapping data (Primary Tumor).

Key Results and Discussion

The evaluation revealed a "Less is More" paradox: in the Lymphography and Breast Cancer domains, removing rare-case rules actually maintained or slightly increased accuracy. This suggests that rare examples often represent noise rather than meaningful exceptions.

Figure 4: Performance of AQ15 across different truncation levels compared to human experts and random choice.

As shown in Figure 4, AQ15 achieved 82% accuracy in Lymphography with only 4 rules, matching human expert estimates (85%). In the Breast Cancer domain, AQ15's 68% accuracy outperformed human specialists (64%). Furthermore, the TRUNC method drastically reduced complexity; in Breast Cancer, the number of logical "selectors" dropped from 160 to 7 without losing predictive power.

However, the "Primary Tumor" domain highlighted the limits of truncation. Due to data sparsity, truncating rules led to a significant drop in accuracy (39% to 29%), indicating that in sparse domains, rare cases may contain vital information.

Conclusion

AQ15 demonstrates that a two-tiered representation—splitting concept meaning between explicit rules and an interpretation algorithm—allows for simpler, more human-readable knowledge bases. While highly effective for many diagnostic tasks, the system's reliance on attribute-based descriptions and its sensitivity to data sparsity in complex domains like Primary Tumor suggest areas for future refinement in automated rule specialization and generalization.

Summary updated

BibTeX
@inproceedings{michalski1986multi,
  author = {Ryszard S. Michalski and Igor Mozetic and Jiarong Hong and Nada Lavrac},
  title = {{The Multi-Purpose Incremental Learning System AQ15 and Its Testing Application to Three Medical Domains}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1986},
}
Figure