Computational Models of Historical Scientific Discoveries
Pat Langley, Lorenzo Magnani, Peter Cheng, Adrian Gordon, Șakir Kocabaș, Derek H. Sleeman · eScholarship (California Digital Library) · 2001
Computational Models of Historical Scientific Discoveries Pat Langley, Institute for the Study of Learning and Expertise Lorenzo Magnani, Department of Philosophy, University of Pavia Peter C.-H. Cheng, School of Psychology, University of Nottingham Adrian Gordon, Department of Computing, University of Northumbria Sakir Kocabas, Space Engineering Department, Istanbul Technical University Derek H. Sleeman, Department of Computing Science, University of Aberdeen The discovery of scientific knowledge is one of the most challenging tasks that confront humans, yet cogni- tive science has made considerable progress toward ex- plaining this activity in terms of familiar cognitive pro- cesses like heuristic search (e.g., Langley et al., 1987). A main research theme relies on selecting historical discov- eries from some discipline, identifying data and knowl- edge available at the time, and implementing a computer program that models the processes that led to the scien- tists’ insights. The literature on computational scientific discovery includes many examples of such studies, but initial work in this tradition had some significant draw- backs, which we address in this symposium. One such limitation was that early research in law dis- covery ignored the influence of domain knowledge in guiding search. For example, Gordon et al. (1994) noted that attempts to fit data from solution chemistry in the late 1700s took into account informal qualitative models like polymerization and dissociation. They have devel- oped Hume, a discovery system that draws on such qual- itative knowledge to direct its search for numeric laws. Hume utilizes this knowledge not only to rediscover laws found early in the history of solution chemistry, but also to explain, at an abstract level, the origins of other rela- tions that scientists proposed and later rejected. Early discovery research also downplayed the role of diagrams, which occupy a central place in many aspects of science. For example, Huygens’ and Wren’s first pre- sentations of momentum conservation took the form of diagrams, suggesting they may have been instrumental in the discovery process. In response, Cheng and Simon (1992) have developed Huygens, a computational model for inductive discovery of this law that uses a psycho- logically plausible diagrammatic approach. The system replicates the discovery by manipulating geometric dia- grams that encode particle collisions and searching for patterns common to those diagrams. The quantitative data given to the system are equivalent to those available at the time of the original discovery. Another challenge concerns the computational model- ing of extended periods in the history of science, rather than isolated events. To this end, Kocabas and Langley (1995) have developed BR4, an account of theory revi- sion in particle physics that checks if the current theory is consistent (explains observed reactions) and complete (forbids unobserved reactions), revises quantum values and posits new particles to maintain consistency, and in- troduces new properties to maintain completeness. BR-4 models, in abstract terms, major developments in par- ticle physics over two decades, including the proposal of baryon and lepton numbers, postulation of the neu- trino, and prediction of numerous reactions. Background knowledge about symmetry and conservation combine with data to constrain the search for an improved the- ory in a manner consistent with the incremental nature of historical discovery. We hope this symposium will encourage additional re- search that extends our ability to model historical scien- tific discoveries in computational terms. References Cheng, P. C.-H. and Simon, H. A. (1992). The right rep- resentation for discovery: Finding the conservation of momentum. In Proceedings of the Ninth International Conference on Machine Learning, pages 62–71, San Mateo, CA. Morgan Kaufmann. Gordon, A., Edwards, P., Sleeman, D., and Kodratoff, Y. (1994). Scientific discovery in a space of structural models. In Proceedings of the Sixteenth Annual Con- ference of the Cognitive Science Society, pages 381– 386, Atlanta. Lawrence Erlbaum. Kocabas, S. and Langley, P. (1995). Integration of research tasks for modeling discoveries in particle physics. In Proceedings of the AAAI Spring Sympo- sium on Systematic Methods of Scientific Discovery, pages 87–92, Stanford, CA. AAAI Press. Langley, P., Simon, H. A., Bradshaw, G. L., and Zytkow, J. M. (1987). Scientific discovery: Computational ex- plorations of the creative processes. MIT Press, Cam- bridge, MA.