Reference: Walker, M. G. How Feasible is Automated Discovery? 1986.
Abstract: There is a widespread belief that a computer can not and will not make an original scientific discovery. It is also commonly believed that discovery is a mystical event that cannot be the result of a planned sequence of steps. Both these beliefs are incorrect. At the same time, many claims about the success of artificial intelligence programs and expert systems are greatly exaggerated. We need to know what has genuinely been learned, and what problems and opportunities exist. That is the subject of this article. Meta-Dendral, the first artificial intelligence program for automated discovery, discovered previously unknown scientific knowledge just over ten years ago. The Bacon program replicated several historical discoveries by following a planned, but general, sequence of steps. It also showed how, in at least one sense, a computer program could formulate new scientific concepts. The RX program discovered medical knowledge not previously known to its builders, using a small medical knowledge base and very general statistical methods. Prospector discovered a mineral deposit missed by human experts in over 60 years of searching. These programs suggest that automated discovery has the potential to be cost-effective. In this article we will examine how each of these programs works, their failings, and potential solutions. Based on this examination, we will look for general patterns and lessons, and suggest how automated discovery might be further pursued.
Notes: Working Paper.