| Abstract
|
We define and implement a model of rationa … We define and implement a model of rational actionfor automated reasoning systems that makes use of flexibleapproximation methods and inexpensive decision-theoretic procedures todetermine how best to solve a problem under bounded computationalresources. The model provides metareasoning techniques which enable areasoning system to balance the costs of increased delays with thebenefits of better results in a decision context. Thedecision-theoretic metareasoning techniques presented can be appliedto a variety of computational tasks. We focus on the use ofinexpensive decision procedures to control complex decision-theoreticreasoning at the base level. The approach extends traditional decisionanalyses to autoepistemic models that represent knowledge aboutproblem solving, in addition to knowledge about distinctions andrelationships in the world. We found that it can be valuable toallocate a portion of costly reasoning resources to metaleveldeliberation about the best way to use additional resources to solve adecision problem.After reviewing principles for applying multiattribute utility theoryto the control of computational procedures, we describe how theseprinciples can be used to control probabilistic reasoning. Inparticular, we shall examine techniques for controlling, at run time,the tradeoff between the complexity of detailed, accurate analyses andthe tractability of less complex, yet less accurate probabilisticinference. Then we review the architecture and functionality of asystem named Protos that embodies the principles for using complexprobabilistic models to make high-stakes decisions under timepressure. We shall study the behavior of Protos on high-stakesdecision problems in medicine. Finally, we move beyond our focus ontime constraints to consider the constraints on decision-theoreticreasoning posed by the cognitive limitations of people seeking insightfrom automated decision systems. ng insightfrom automated decision systems.
|
