2001) Computational discovery of communicable knowledge: Symposium report

versity. Fifteen speakers reviewed recent advances in computational ap-proaches to scientific discovery, focusing on their discovery tasks and the generated knowledge, rather than on the discovery algorithms them-selves. Despite considerable variety in both tasks and methods, the talks were unified by a concern with the discovery of knowledge cast in for-malisms used to communicate among scientists and engineers. Computational research on scientific discovery has a long history within both artificial intelligence and cognitive science. Early efforts focused on reconstruct-ing episodes from the history of science, but the past decade has seen similar techniques produce a variety of new scientific discoveries, many of them leading to publications in the relevant scientific literatures. Work in this paradigm has emphasized formalisms used to communicate among scientists, including numeric equations, structural models, and reaction pathways. However, in recent years, research on data mining and knowledge discovery has produced another paradigm. Even when applied to scientific domains, this framework employs formalisms developed by artificial intelligence researchers themselves, such as decision trees, rule sets, and Bayesian networks. Although such methods can produce predictive models that are highly accurate, their outputs are not stated in terms familiar to scientists, and thus typically are not very communicable. To highlight this distinction, Pat Langley organized the Symposium on Com-putational Discovery of Communicable Knowledge, which took place at Stanfor