Abstract:
Formal decision-making quality is affected by the capabilities of the decision- making tools and the availability of reusable decision knowledge. Decision graphs are a powerful decision modeling framework. Many decision graph elements are probabilistic, and each requires a Local Probability Distribution (LPD) specifying the probabilities of each state in that element. State-of-the-art implementations have first-order expressivity, where the computer can automatically create decision graphs with a varying number of entities (e.g. people, decisions, alternatives, etc.). Varying the entities creates different model configurations. For every possible model configuration, the modeling tool must define the LPD, using information provided by the modeler. Existing first-order expressive decision graph implementations have a critical limitation arising from a lack of re-search on the full range of capabilities required by a modeling tool to create these LPD. This lack restricts modeling of important types of decision problems. Compounding this limitation is a lack of decision knowledge reuse capabilities, reducing decision modelers’ ability to learn from other modelers’ experiences on similar decision problems, in-creasing modeling time and the possibility of an incomplete model. The research focused on eliminating modeling tool limitations and creating a knowledge reuse capability. First, it developed a new approach, dependency modeling, that identified a set of LPD behaviors that provide a robust capability to any first-order expressive modeling tool and includes ten new LPD development language capabilities that significantly extend the range of problems that can be modeled. Second, dependency modeling uncovered eight design patterns useful for modeling in any probabilistic first-order expressive framework. Third, it refined and integrated four factors that differentiate decision problems, identifying information reuse requirements and uncovering additional modeling tool needs. Fourth, it addressed the additional needs uncovered, developing algorithms to address decision problem asymmetry, developing a more efficient approach to model context variation, and defining modeling approaches for decision problems with varying entity counts. Fifth, it created decision templates as a decision knowledge reuse tool. An initial evaluation provided preliminary data that the template could reduce model development time by 50%, with 75% of participants assessing the template as useful or very useful in completing the experimental task.
