Goal-directed qualitative reasoning with partial states

This research explores the representational and computational complexities of qualitative reasoning about time-varying behavior. Traditional techniques employ qualitative simulation (QS) to compute envisionments (i.e. state-transition graphs) representing all possible behaviors. Unfortunately, QS exhaustively case-splits on all choices, regardless of specific task goals. It reasons with completely described states and explores every (ambiguous) future of each.In this thesis we introduce a new representation, called sufficient discriminatory envisionments (SUDE's), which addresses these problems. SUDE's discriminate the possible behavior space by whether the goal is possible, impossible, or inevitable from each state in that space. Our techniques for generating SUDE's strive to reason with the smallest state descriptions which are sufficient for making these discriminations.We present algorithms for generating SUDE's via a two-stage process. First, exhaustive regression sketches the space of possible paths between the initial and goal states. Second, we qualify these possible paths, identifying conditions under which the goal is impossible or inevitable and finding all possible transitions between these paths.We formulate Nature's regression operators in terms of minimal chunks of causality, exploiting the causal, compositional nature of Qualitative Process Theory models. We integrate continuity-based and minimality-based theories of change to support discontinuous change due to actions and modelling simplifications.We discuss our implementation of these techniques and our test examples in three domains, which we call ball-world, tank-world, and kitchen-world.U of I OnlyETDs are only available to UIUC Users without author permissio