Dynamic application of problem solving strategies : dependency-based flow control

Abstract

While humans may solve problems by applying any one of a number of different problem solving strategies, computerized problem solving is typically brittle, limited in the number of available strategies and ways of combining them to solve a problem. In this thesis, I present a method to flexibly select and combine problem solving strategies by using a constraint-propagation network, informed by higher-order knowledge about goals and what is known, to selectively control the activity of underlying problem solvers. Knowledge within each problem solver as well as the constraint-propagation network are represented as a network of explicit propositions, each described with respect to five interrelated axes of concrete and abstract knowledge about each proposition. Knowledge within each axis is supported by a set of dependencies that allow for both the adjustment of belief based on modifying supports for solutions and the production of justifications of that belief. I show that this method may be used to solve a variety of real-world problems and provide meaningful justifications for solutions to these problems, including decision-making based on numerical evaluation of risk and the evaluation of whether or not a document may be legally sent to a recipient in accordance with a policy controlling its dissemination.