Discovering the Wave Theory of Sound: Inductive Inference in the Context of Problem Solving

Paul Thagard, Keith J. HolyoakIJCAI 1985

Introduction: Beyond Curve Fitting

Early artificial intelligence research into scientific discovery primarily focused on identifying empirical laws. Systems like the BACON series were adept at finding mathematical patterns in data—such as Kepler’s third law—by fitting curves to observed variables. However, the most significant milestones in science often involve the creation of theories rather than just laws. Theories do more than describe data; they explain it by postulating hidden, non-observable entities like atoms, gravity, or waves.

This paper introduces PI (Processes of Induction), a computational system designed to address a fundamental question: How can an AI generate a brand-new theoretical concept to solve a specific explanatory problem? The authors argue that scientific discovery is not a standalone "eureka" moment but a byproduct of active, goal-directed problem solving triggered by the need to explain puzzling phenomena.

The PI Framework: A Hybrid Architecture

The PI system utilizes a hybrid architecture that combines two classic AI structures to represent knowledge and drive inference:

  1. Production Rules: "If-then" statements that govern the system's actions and logical transitions.
  2. Frames: Concept clusters (e.g., "Water" or "Wave") that store properties, expectations, and variability judgments.

Unlike previous systems where activation of concepts was automatic and exhaustive, PI employs a goal-directed "spreading activation" model. In this model, the system only activates concepts and rules relevant to the specific problem it is trying to solve, mimicking the focused nature of human scientific inquiry.

The Four Pillars of Induction

The core of the PI system is its ability to perform four distinct types of induction, ranging from simple generalizations to complex conceptual innovation:

  • Instance-Based Generalization: The system learns "All A are B" from specific examples. Crucially, it accounts for the variability of categories; for instance, it recognizes that birds vary in color more than metals vary in their ability to conduct electricity.
  • Condition-Based Generalization: This involves merging specific rules into broader ones. For example, observing that multiple types of fast-food burgers are greasy allows the system to generalize the property to the broader category of "burgers."
  • Abduction: This is the process of generating a hypothesis to explain a fact. If the system knows that sound reflects and that waves reflect, it may use abduction to hypothesize that "sound is a wave."
  • Conceptual Combination: This is the system's most novel feature. It creates a new concept (e.g., "Sound-Wave") by merging two existing concepts and resolving their inherent contradictions. If water waves are known to spread in a plane while sound is known to spread spherically, the system must use "top-down" variability judgments to decide which property the new hybrid concept should inherit.

Case Study: Reconstructing the Wave Theory of Sound

To test the system, the authors provided PI with hand-coded knowledge bases representing the state of scientific understanding in antiquity, drawing from the writings of the Roman architect Vitruvius and various Greek philosophers. The system successfully simulated two distinct historical paths to the discovery of the wave theory of sound:

The Vitruvius Path

This path was triggered by the goal of explaining echoes (reflection). PI generalized the rule that "all waves reflect" and used abduction to hypothesize that sound might be a wave. Through conceptual combination, it synthesized the "sound-wave" concept, successfully resolving the conflict between the planar propagation of water waves and the spherical propagation of sound.

The Greek Path

This path was rooted in acoustics and the study of stringed instruments. PI reached the wave theory through a chain of associations between sound, musical pitch, the physical vibration of strings, and the eventual connection to wave-like motion.

Theory Evaluation and Rejection

PI does not blindly accept every hypothesis it generates. The system includes mechanisms for theory rejection based on explanatory failure. For example, the system "discovered" the particle theory of sound (associated with Democritus) but ultimately rejected it. The particle theory failed to solve a critical explanatory problem: how two sounds can pass through each other without their constituent particles colliding and disrupting the signal. This demonstrates that PI’s discovery process is deeply integrated with its ability to evaluate a theory's problem-solving utility.

Conclusion and Limitations

The PI system demonstrates that a unified computational architecture can handle both mundane problem-solving and high-level theoretical innovation. However, the authors note several limitations. The system's "discoveries" are heavily dependent on the specific rules and frames pre-loaded into its knowledge base. Furthermore, while spreading activation limits the search space, future iterations would require more sophisticated methods for evaluating which new rules are most efficient for problem-solving. Despite these challenges, PI provides a robust model for how non-observable theoretical entities can be induced through the process of resolving explanatory gaps.

Summary updated

BibTeX
@inproceedings{thagard1985discovering,
  author = {Paul Thagard and Keith J. Holyoak},
  title = {{Discovering the Wave Theory of Sound: Inductive Inference in the Context of Problem Solving}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1985},
}
Figure