Creating generalized procedures by demonstration

We present key components of the ARANDA system for constructing algorithmic, generalized plans for solving broad classes of problems by generalizing solutions of simple problem instances. These generalized plans are powerful enough to solve problem instances varying not only in terms of object properties, but also in the numbers of objects themselves, which may be unbounded. This is achieved by using loops to manipulate the potentially unbounded numbers of objects; in contrast to other approaches, these loops are guaranteed to make measurable progress, as well as to terminate. We also present methods for efficiently determining the class of problem instances which can be solved by a computed generalization.