An integrated approach to dynamic decision making under uncertainty

Decision making is often complicated by the dynamic and uncertain information involved. This work unifies and generalizes the major approaches to modeling and solving a subclass of such decision problems. The relevant problem characteristics include discrete problem parameters, separable optimality functions, and sequential decisions made in stages. The relevant approaches include semi-Markov decision processes, dynamic decision modeling, and decision-theoretic planning. An analysis of current decision frameworks establishes a unifying task definition and a common vocabulary; the exercise also identifies the trade-off between model transparency and solution efficiency as their most significant limitation. Insights gained from the analysis lead to a new methodology for dynamic decision making under uncertainty. The central ideas involved are multiple perspective reasoning and incremental language extension. Multiple perspective reasoning supports different information visualization formats for different aspects of dynamic decision modeling. Incremental language extension promotes the use of translators to enhance language ontology and facilitate practical development. DynaMoL is a language design that adopts the proposed paradigm; it differentiates inferential and representational support for the modeling task from the solution or computation task. The dynamic decision grammar defines an extensible decision ontology and supports problem specification with multiple interfaces. The graphical presentation convention governs parameter visualization in multiple perspectives. The mathematical representation as semi-Markov decision process facilitates formal model analysis and admits multiple solution methods. A general translation technique is devised for the different perspectives and representations of the decision factors and constraints. DynaMoL is evaluated on a prototype implementation, via a comprehensive case study in medicine. The case study is based on an actual treatment planning decision for a patient with atrial fibrillation. The problems addressed include both long-term discrimination and short-term optimization of different treatment strategies. The results demonstrate practical promise of the framework. Thesis supervisor: Peter Szolovits Title: Professor of Computer Science and Engineering