Preliminary Design of an Empirical Discovery Agent

We approach the problem of automated knowledge discovery from an empirical perspective. We are developing an artiicial agent, called the Scientist's Empirical Assistant (SEA), that discovers the structure of the world by formulating hypotheses and testing them through controlled experimentation. SEA is a mixed-initiative system which supports a human scientist in common scientiic activities, such as reasoning about a problem, formulating and testing hypotheses, running experiments, and keeping records of the results. SEA uses planning and heuristic reasoning to search a space of hypotheses for those which accurately reeect the observed behavior of the world. It relies heavily on the user's expert knowledge of the domain to guide the planner toward the most appropriate course of action. We discuss the design of SEA through the presentation of an annotated example. The example application is a non-trivial, open scientiic problem from a real-world domain. While we cannot claim that SEA achieves any novel discoveries, its results are suprising given the simplicity of the design.