OBJECTIONS TO DISCOVERY BY COMPUTER

This chapter discusses the objections to the discovery of artificial intelligence programs by computer. It presents the objections that apply to some forms of computer discovery or computer-aided search for theories: (1) people have various kinds of special knowledge that computers do not have, and that knowledge makes people better at discovery than computers can possibly be; (2) discovery is a form of inference and inference should proceed in accordance with the requirements of Bayesian statistics and Bayesian epistemology, but computer programs that aid in scientific discovery often do not work on strict Bayesian principles; and (3) offering computer programs that are intended to aid in the process of discovery is like playing with fire. Artificial intelligence programs for discovering statistical models are really carrying out decision procedures, but they do not act like rational Bayesian agents. Even if a program has carried out Bayesian calculations, whatever prior probability distribution and utilities the program uses may not be shared by human researchers. From a Bayesian perspective, the TETRAD program is a device for investigating a part of the catch-all hypothesis and for locating within it specific alternatives that give the evidence a reasonable likelihood and that have the virtues of simplicity and explanatory power.