Browse Theses

We define and implement a model of rational action for automated reasoning systems that makes use of flexible approximation methods and inexpensive decision-theoretic procedures to determine how best to solve a problem under bounded computational resources. The model provides metareasoning techniques which enable a reasoning system to balance the costs of increased delays with the benefits of better results in a decision context. The decision-theoretic metareasoning techniques can be applied to a variety of computational tasks. We focus on the use of metalevel decision procedures to control complex probabilistic reasoning at the base level. The approach extends traditional decision analyses to autoepistemic models that represent knowledge about problem solving, in addition to knowledge about distinctions and relationships in the world. We found that it can be valuable to allocate a portion of costly reasoning resources to metalevel deliberation about the best way to use additional resources to solve a decision problem.After reviewing principles for applying multiattribute utility theory to the control of basic computational procedures, we describe how these principles can be used to control probabilistic reasoning. In particular, we examine techniques for controlling, at run time, the tradeoff between the complexity of detailed, accurate analyses and the tractability of less complex, yet less accurate probabilistic inference. We present the architecture and functionality of a system named Protos that embodies these principles for making high-stakes decisions under time pressure. We study the behavior of Protos on decision problems in critical-care medicine. Finally, we move beyond our focus on time constraints to consider the constraints on decision-theoretic reasoning posed by the cognitive limitations of people seeking insight from automated decision systems.