Search for Regularity: Four Aspects of Scientific Discovery

摘要

Scientific discovery is a complex activity involving many different components. Our interest in discovery has led us to construct four artificial intelligence systems that address different facets of this process. BACON.6 focuses on discovering empirical laws that summarize numerical data. This program searches a space of data and a space of numerical laws, and includes methods for postulating intrinsic properties and noting common divisors. GLAUBER is concerned with discovering laws of qualitative structure, such as the hypothesis that acids react with alkalis to form salts. It searches the space of qualitative laws, using evaluation functions to focus attention on laws covering the greatest number of observed facts. STAHL attempts to determine the components of substances involved in reactions, and has been used to model the reasoning that led to the phlogiston theory. This system searches through the space of componential models, using heuristics to make plausible inferences. The final system, DALTON, is concerned with formulating structural models of chemical reactions. It searches the space of possible models, considering simple models before more complex ones and using a conservation assumption to constrain possibilities. While each of these discovery systems is interesting in its own right, we are also exploring ways in which the systems can interact to help direct each other's search processes. (Author)