Automated Discovery a Fusion of Multidisciplinary Principles.pdfVIP

Automated Discovery: a Fusion of Multidisciplinary Principles Jan M. _Zytkow Computer Science Department, UNC Charlotte, Charlotte, N.C. 28223 and Institute of Computer Science, Polish Academy of Sciences zytkow@uncc.edu After two decades of research on automated discovery (1995; 1998; 1999; Chaudhuri Madigan, 1999; Edwards, 1993; Komorowski Zytkow, 1998; Lan- gley, Simon, Bradshaw, Zytkow, 1987; Piatetsky-Shapiro Frawley, 1991; Shen, 1993; Shrager Langley; Simon, Valdes-Perez Sleeman, 1997; Zytkow, 1992, 1993, 1997), it is worthwhile to summarize the foundations for discovery systems by a set of principles. We propose a number of such principles and we discuss the ways in which dierent principles can be used together to explain discovery systems and guide their construction. Automated discovery is closely linked to natural sciences, logic, philosophy of science and theory of knowledge, articial intelligence, statistics, and machine learning. Knowledge discovery tools use a creative combination of knowledge and techniques from the contributing areas, adding its own extra value, which we emphasize in several principles. 1 What is a discovery We start by clarifying the notion of discovery that applies to automated agents. A person who is rst to propose and justify a new piece of knowledge K is con- sidered the discoverer of K. Being the rst means acting autonomously, without reliance on external authority, because there was none at the time when the discovery has been made, or the discovery contradicted the accepted beliefs. Machine discoverers should be eventually held to the same standards. Abso- lute novelty is important, but a weaker criterion of novelty is useful in system construction: Agent A discovered knowledge K i A acquired K without the use of any knowledge source that knows K. This denition calls for cognitive autonomy of agent A. It requires only that K is novel to the agent, but does not have to be found for the rst time in the human histor