Empirical Methods in AI

Empirical methods have been successful in recent years. Indeed, as Henry Kautz reminded the workshop participants, in the last year alone, the New York Times has reported two major empirical successes: (1) DEEP BLUE’s defeat of Kasparov and (2) the computer-generated proof of an open problem in Robbins algebra. Pandurang Nayak (NASA Ames) described another highly publicized success, the diagnosis system for the Deep Space One spacecraft, which is based on a highly optimized satisfiability procedure. Although deciding satisfiability is intractable in general, this system generates plans in practice in essentially constant time for each step. It comes as quite a surprise to hear about real-time satisfiability testing. Henry Kautz listed several reasons for the success of empirical methods. First, empirical studies are often an integral part of AI because systems can be too complex or messy for theory. Second, theory is often too crude to provide useful insight. For example, a problem might be exponential in the worst case but tractable in practice. Third, some questions are purely empirical. As Pedro Meseguer (IIIA, CSIC, Spain) pointed out during one of the panels, two search algorithms ■ In the last few years, we have witnessed a major growth in the use of empirical methods in AI. In part, this growth has arisen from the availability of fast networked computers that allow certain problems of a practical size to be tackled for the first time. There is also a growing realization that results obtained empirically are no less valuable than theoretical results. Experiments can, for example, offer solutions to problems that have defeated a theoretical attack and provide insights that are not possible from a purely theoretical analysis. I identify some of the emerging trends in this area by describing a recent workshop that brought together researchers using empirical methods as far apart as robotics and knowledge-based systems.