Learning about Momentum Conservation

Jude W. ShavlikIJCAI 1985

Published in the mid-1980s, this research represents a significant intersection of classical Artificial Intelligence and physics education. It introduces Physics 101, a system designed to bridge the gap between fundamental physical laws (such as Newton’s) and high-level problem-solving shortcuts (such as the Conservation of Momentum) using a technique known as Explanation-Based Learning (EBL).

Overview

Physics 101 is an AI system that "discovers" general physical principles by analyzing human-provided solutions to specific problems. By grounding its reasoning in Newton’s laws and symbolic calculus, it generalizes a simple one-dimensional collision solution into a universal formula for momentum change in complex three-dimensional systems.

The Problem of Physical Learning

In the 1980s, AI systems for physics generally fell into two categories: those that performed rote algebraic manipulation and those that attempted to "learn" laws from massive datasets. Neither approach truly captured the nuance of human learning.

When a student learns about the conservation of momentum, they are ideally understanding it as a logical consequence of Newton’s laws rather than just memorizing a formula. The challenge addressed here is: How can a machine use "deep" domain knowledge (fundamental laws) to justify and then generalize a specific problem-solving "shortcut" provided by a human?

Methodology: Explanation-Based Learning

The system utilizes Explanation-Based Learning (EBL). Unlike modern deep learning, which requires thousands of examples, EBL allows a system to learn a general rule from a single example by "explaining" why that specific example worked.

The Learning Loop

The Physics 101 learning process flowchart, illustrating the cycle from problem failure to knowledge update. Source, p. 1

The Physics 101 learning process flowchart, illustrating the cycle from problem failure to knowledge update.
  1. Initial Knowledge: Physics 101 begins with Newton’s 2nd Law ($F=ma$), Newton’s 3rd Law (action/reaction), and the ability to perform symbolic calculus (differentiation and integration).
  2. Verification and Generalization:
    • The system is presented with a problem it cannot solve (e.g., a collision where specific impact forces are unknown).
  3. A human provides a solution strategy (e.g., "use conservation of momentum").
  4. The system verifies the solution using calculus to prove that internal forces between objects cancel each other out due to Newton’s 3rd Law.
  5. Once verified, it generalizes the result, realizing this cancellation occurs regardless of the number of objects or dimensions involved.

Experimental Scenarios

As was typical for AI research of this era, the paper utilizes specific scenarios rather than large-scale datasets to demonstrate capability.

A training scenario depicting a one-dimensional, two-body collision in free space.
  • Training Scenario: A 1D, two-body collision in free space with no external forces. This serves as the basis for the system to derive the principle of conservation.
  • Test Scenario: A 3D, three-body collision occurring in the presence of an external force. In this case, momentum is not conserved, but the system's learned general formula still correctly predicts the outcome.

Key Insights and Results

  • The Power of "Deep" Models: The system does not blindly accept human input. It must derive the formula from first principles. If a human provides an incorrect step, the system rejects it.
  • Generalization through Cancellation: The core discovery for the system is that internal forces ($F_{12}$ and $F_{21}$) always sum to zero. This allows the system to ignore complex, unknown interaction forces during a collision.
  • Successful Derivation: Physics 101 successfully derived a general principle: The total momentum of a system changes at a rate determined by the integral of the sum of external forces. After learning from a simple 1D problem, the system was able to solve complex 3D problems involving external forces—a task it was mathematically incapable of performing before training.

Limitations and Future Outlook

The system’s learning is strictly limited by its symbolic math capabilities; if the required calculus is too complex, the system cannot learn the underlying principle. Furthermore, it is not fully autonomous, as it requires a human to provide the initial strategy.

However, this research suggests powerful applications for Intelligent Tutoring Systems (ITS). By understanding the derivation of physical laws, an AI could detect exactly where a student’s conceptual understanding has failed, such as incorrectly assuming momentum is conserved in the presence of gravity.

Summary updated

BibTeX
@inproceedings{shavlik1985learning,
  author = {Jude W. Shavlik},
  title = {{Learning about Momentum Conservation}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1985},
}
Figure