| Abstract
|
I define and implement a model of rational … I define and implement a model of rational action for automated reasoningsystems that makes use of flexible approximation methods anddecision-theoretic procedures to determine how best to solve a problemunder bounded computational resources. The model provides a perspective onthe use of metareasoning techniques to balance the costs of increaseddelays with the benefits of better results in a decision context. I focuson the use of inexpensive real-time analyses to control the allocation ofcomputational resources in complex decision-theoretic reasoning. Theapproach extends traditional decision analyses to autoepistemic modelsthat represent knowledge about problem solving, in addition to knowledgeabout distinctions and relationships in the world. To investigate the use of decision analysis for controlling computation, Iconstructed a computer program named Protos. Protos uses information aboutthe progress of problem solving to identify the ideal time to haltcomputation and take action in the world. Protos' metareasoner controlsthe precision of probabilities inferred from complex network models thatrepresent domain-specific expertise about uncertain relationships amongobservations and hypotheses. I found that it can be valuable to allocate aportion of costly reasoning resources to deliberate about the best way tosolve a decision problem. In addition to serving as a testbed forexploring the value of metareasoning, I made use of Protos to examine theintegration of reflex and deliberative analyses and the construction oftime-dependent utility models from observations. After discussing principles for applying multiattribute utility theory tothe control of basic computational procedures, I describe how theseprinciples can be used to control probabilistic reasoning. In particular,I present techniques for controlling, at run time, the tradeoff betweenthe complexity of detailed, accurate analyses and the tractability of lesscomplex, yet less accurate probabilistic inference. Then, I describe thearchitecture and functionality of Protos and review the system's behavioron high-stakes decision problems in medicine. Finally, I move beyond theconsideration of time constraints to investigate the constraints ondecision-theoretic reasoning posed by the cognitive limitations of peopleseeking insight from automated decision systems. g insight from automated decision systems.
|
