Generating Predictions to Aid the Scientific Discovery Process

Randy JonesAAAI 1986

Introduction

Presented at AAAI-86, NGLAUBER is an incremental machine learning system designed to discover qualitative scientific laws. Evolving from the earlier GLAUBER system, NGLAUBER shifts from passive data analysis to an active model of discovery by generating testable predictions. This approach mirrors the iterative nature of the scientific method, where theories guide the collection of new data.

Beyond Passive Data Analysis

In the mid-1980s, most scientific discovery systems were batch-oriented, requiring a complete dataset before analysis could begin. This "passive" approach failed to capture how scientists actually work: data arrives incrementally, and current hypotheses dictate which experiments to perform next.

NGLAUBER addresses this by focusing on incremental conceptual clustering. Unlike contemporary algorithms that grouped objects by shared attributes (e.g., color), NGLAUBER clusters objects based on their relational properties. For instance, it might group substances not because they look alike, but because they react similarly with a third substance.

The NGLAUBER Framework

NGLAUBER operates through four recursive mechanisms that allow it to build and refine its understanding of a domain:

  1. Introduction: Upon observing similar behaviors (e.g., two chemicals both tasting sour), the system creates a new class (e.g., "acids") and reformulates its knowledge base using these labels.
  2. Prediction: This is the core "aha" engine. When the system identifies a property in some members of a class, it hypothesizes that the property applies to all members.
  3. Prediction Satisfaction: As new data is ingested, the system validates its existing predictions. Success strengthens the underlying hypothesis.
  4. Denial: If a prediction is contradicted by new evidence, the system prunes the entire "prediction group" or hypothesis. This prevents the accumulation of false theories.
Hierarchy of logical statements ranging from most specific (existential) to most general (universal).

The "Prediction" mechanism effectively navigates a logical space, as shown in the figure above. By transforming existential statements ($\exists$) into universal ones ($\forall$), the system moves from specific observations toward the "most general" laws possible.

Case Study: 18th-Century Chemistry

The system was evaluated using a benchmark of 18th-century chemical knowledge. Given 12 facts about chemical reactions and qualitative properties like taste, NGLAUBER successfully reconstructed the classes of "acids," "alkalis," and "salts." It also derived the general law that the reaction of any acid with any alkali produces a salt.

Notably, the system remained efficient despite a high initial rate of false predictions (60% to 85%). The "Denial" mechanism acted as a rigorous filter, pruning incorrect hypotheses and leaving only historically accurate qualitative laws.

Discussion and Future Directions

NGLAUBER demonstrates that scientific categories are often better defined by interactions than by static attributes. By treating predictions as a "sieve," the system maintains a maximally general description of the world without losing the ability to handle specific details.

However, the system faces limitations, particularly its intolerance to noise. A single outlier can cause the system to discard a valid hypothesis, suggesting a need for probabilistic reasoning in future iterations. Additionally, while the current system produces flat partitions, future work aims to incorporate hierarchical clustering to discover more complex structures, such as the Periodic Table.

By shifting AI from a passive data-cruncher to an active participant that proposes its own "experiments," NGLAUBER provides a compelling model for automated scientific discovery.

Summary updated

BibTeX
@inproceedings{jones1986generating,
  author = {Randy Jones},
  title = {{Generating Predictions to Aid the Scientific Discovery Process}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1986},
}
Figure