📄 BACON: A Production System That Discovers Empirical Laws

Summary

BACON asks whether a program can rediscover simple empirical laws like those of early physicists, following scientific-reasoning systems such as DENDRAL, meta-DENDRAL, MYCIN and AM. Its testbed is an artificial two-dimensional world of labeled objects whose primitive attributes (x-coordinate, color, velocity, etc.) vary across objects and time under laws chosen by the programmer. Written in the production-system language OPS, BACON keeps observations as attribute-object-time-value tuples and runs a few general regularity detectors over them in parallel. A constancy triggers a tested generalization over time, objects, or both. A steady trend yields a new rate-of-change attribute, such as acceleration derived from velocity. An inverse relation between two attributes yields their product. Repeating this eventually builds a composite attribute with a constant value, which is the system's version of the law. Confirmed generalizations become prediction rules, and counterexamples produce qualifications instead of rejection. The paper is an early demonstration that regularity detection combined with the construction of new terms can automate law discovery.

Key points

• Implemented in OPS (Forgy & McDermott, 1976). The production-system format lets a small set of generic "regularity detectors" scan working-memory data, stored as attribute-object-time-value 4-tuples, in parallel.
• There are three core heuristics. Constant values lead to generalization. Monotonic change over time leads to a derivative-style attribute. Opposite-direction co-variation of two attributes leads to a product attribute.
• An example hidden law is that the third derivative of each object's distance from the universe's center is constant, with a different constant for each object.
• Hypotheses are never discarded. They are kept in production memory and qualified with exceptions, or with exceptions plus upper and lower bounds for generalizations over time. Data gathering starts with random attribute sampling and becomes guided once regularities are detected.
• The author lists three limitations: the system cannot form conditional laws, it relies on built-in knowledge that time and objects are the dimensions to generalize over, and it cannot handle noisy data. The author says the noise problem could be fixed by a minor change to OPS numeric pattern matching.