The Ubiquity of Discovery: Computers and Thought

IJCAI 1977

TL;DR

This paper argues that "intelligence," particularly in creative and problem-solving domains, can be modeled as a "heuristic rule-guided search." The authors propose that by embodying this theory in computer programs, we can gain insights into human cognition and even automate aspects of scientific discovery and invention.

What Problem Does This Paper Address?

The paper tackles the fundamental question of how intelligent behavior, especially creative discovery and problem-solving, arises. While scientific models in physics can be highly predictive and quantitative, understanding human intelligence, which is complex and less easily isolated, has historically relied on descriptive prose or metaphorical explanations. The authors aim to bridge this gap by proposing a computational model that can explain and even replicate aspects of human intelligence, moving beyond mere description to a more rigorous, testable theory.

Methods

The core methodology presented is the "heuristic rule-guided search" theory. This approach views intelligent activity as an exploration through a vast space of possibilities, guided by a large collection of informal "rules of thumb" or heuristics. The paper outlines a paradigm for Artificial Intelligence (AI) research:

  1. Identify a cognitive activity: Choose a complex human task (e.g., proving theorems, playing chess, scientific discovery).
  2. Formulate a theory: Develop hypotheses about the underlying information processing mechanisms.
  3. Build a computational model: Implement the theory in a computer program.
  4. Experiment and refine: Observe the program's behavior and use experiments to test and improve the theory.

The paper then details specific AI programs (models) that embody this theory, including:

  • Logic Theorist (LT) and General Problem Solver (GPS): Early attempts at general problem-solving using heuristics.
  • DENDRAL: An expert system for scientific problem-solving in organic chemistry, which incorporated domain-specific heuristics extracted from human experts.
  • AM (A Program That Discovers Mathematics): A program designed for scientific invention, specifically discovering new mathematical concepts and relationships. AM uses a large collection of general heuristics to guide its exploration and judge the "interestingness" of its discoveries.

Datasets

Not discussed in the paper. The paper focuses on the process of discovery and problem-solving, using abstract concepts and simulated tasks rather than specific, pre-defined datasets. The "data" AM works with are mathematical concepts it generates itself.

Key Insights and Ideas

  • Intelligence as Heuristic Search: The central idea is that much of what we perceive as intelligence, from everyday tasks to scientific breakthroughs, can be understood as a search process guided by informal rules of thumb (heuristics).
  • Ubiquity of Heuristics: These heuristic rules are not confined to highly specialized scientific endeavors but are also crucial for mundane activities like cutting cheese or finding one's way around a city.
  • The Power of "Interestingness": The paper proposes that judging the "interestingness" of a concept or discovery is a critical component of intelligence and can itself be modeled using heuristics. This is demonstrated with AM's evaluation of mathematical concepts.
  • Bridging Everyday and Scientific Cognition: The paper shows how the same heuristic principles can explain both everyday problem-solving and invention (e.g., inventing a wire cheese cutter) and scientific discovery (e.g., discovering prime numbers).
  • Model-Building as Scientific Inquiry: The AI paradigm itself is presented as a scientific method for studying intelligence, where computational models serve as testable hypotheses.

Key Results

  • AM's Successes: AM, guided by its heuristics, successfully discovered a significant number of new mathematical concepts and relationships, including primes, highly composite numbers, and Goldbach's conjecture. It demonstrated the ability to generate new concepts and add them to its agenda for further study.
  • Heuristic Generalization: The paper highlights how heuristics used in everyday invention (like heuristic #2: "Consider what variables affect the success/failure...") can be directly applied to scientific discovery (leading to the concept of prime numbers).
  • De-mystification of Creativity: The work suggests that aspects of creativity, such as illumination and incubation, can be understood and potentially mechanized through heuristic rule-guided search.

Limitations and Future Work

  • Heuristic Weakness: As AM ran longer, its general set-theoretic heuristics became too weak to guide its exploration effectively. The paper acknowledges the need for "meta-level rules" (heuristics that create or modify other heuristics) to address this.
  • Need for Meta-Heuristics: The authors are developing EURISKO, a program designed to discover new heuristic rules, recognizing that even professional scientists struggle with formulating and recognizing heuristics.
  • Domain Specificity vs. Generality: The paper touches on the trade-off between general heuristics (like those in GPS) and powerful, domain-specific heuristics (like those in DENDRAL). Future work aims to balance these.
  • Representation: While de-emphasized, the paper notes that the representation of knowledge within AI systems is a crucial, ongoing challenge.

Summary updated

BibTeX
@inproceedings{1977ubiquity,
  title = {{The Ubiquity of Discovery: Computers and Thought}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1977},
}
Figure