Constructive interpretation of human-generated exceptions during plan execution
Carol A. Broverman · Scholarworks (University of Massachusetts Amherst) · 1991
In realistic settings, several factors limit the utility of planners based on the classical nonlinear hierarchical paradigm. The need for human supervision, the inherent incompleteness and incorrectness of models of complex domains, the dynamic nature of the real world, and the frequency of unexpected occurrences while carrying out a calculated plan all imply the inevitability of run-time discrepancies. Previous approaches have viewed deviations from expectations as destructive events, devoid of any valid semantic basis. In contrast, this thesis promotes a constructive approach. We adopt the assumption that the behavior of human agents who interact with the planning system is purposeful, even when it may be inconsistent with system expectations. We claim that the actions of human agents are often misinterpreted as failures, due to gaps in the domain model. This thesis focuses on an interactive planning framework, and describes an approach that intelligently resolves detected inconsistencies resulting from the action of human agents. Through a process of controlled explanation, unusual occurrences are incorporated as valid contributory events, thus avoiding extensive replanning. Explanations are based on a theory of rationales, encoded as plausible inference rules and operationalized for use as a problem detection and correction mechanism. Initial points of failure are categorized according to a taxonomy of exceptions, and the relevance of an exception to the current plan context and domain model is established. Amendments are then made to the partial plan and/or the domain model to restore consistency. A robust planning architecture results that resolves unanticipated contingencies in a constructive fashion whenever possible. In addition, exception handling is used as an opportunistic entry point through which additional knowledge acquisition and refinement takes places. New knowledge acquired during the handling of exceptions produces an augmented knowledge base and improved subsequent system performance. A Common-Lisp prototype demonstrating the approach described in this thesis is implemented on a TI-Explorer. User studies support the adequacy and coverage of our exception taxonomy and rationales.