Data-Driven Approaches to Empirical Discovery

In the last decade a few artificial intelligence researchers have turned their attention to a domain often considered the realm of genius - scientific discovery. The vast majority of this work has focused on empirical discovery, and much of the effort has been concerned with the discovery of numeric laws. This paper traces one evolutionary chain of research on discovery in particular the development of data-driven heuristic methods relating to numeric discovery. The authors examine four systems - Gerwins function induction system, Langley, Bradshaw, and Simons BACON, Zytkows FAHRENHEIT, and Nordhausen and Langleys IDS - and describe how each program introduces abilities lacking in earlier systems. The conceptual advances involve three different but interrelated aspects of discovery the form of laws and theoretical terms discovered the ability to determine the scope and context of laws and the ability to design experiments. This document evaluates each of the systems, but focuses on their theoretical contributions rather than on reporting their behavior in specific domains. It closes the paper by reviewing the work on machine discovery from the views of the history and philosophy of science. KR