This lecture will provide an overview of computational approaches to discovering scientific knowledge. It will characterize discovery as heuristic search through a space of laws or models and distinguish it from work on data mining. Discussion will cover six different types of scientific discovery and review systems that address each of them, as well as attempts to unify and integrate them. Examples will include rediscoveries of laws and models from the history of science and new discoveries that have appeared in the scientific literature. The lecture will close with a list of challenges that future research should address.
This lecture discusses the induction of quantitative process models from background knowledge and multivariate time series. A model contains a set of processes, each with one or more differential equations and a set of associated parameters. The effects of processes are additive, so one can compile any such model into a set of differential equations and simulate its behavior over time. Background knowledge includes generic processes that specify types of variables, forms of differential equation, and ranges for parameter values. Inductive process modeling involves constrained search through a discrete space of model structures and a continuous space of model parameters. The lecture will cover different approaches to discovering process model, with examples taken from population dynamics and chemical reaction networks. It will close with a discussion of limitations and directions for future research.