Decision‐theoretic control of inference for time‐critical applications

Systems that operate successfully in the real world have to coordinate their behavior with the behavior of processes outside their immediate control. Since the decision making required for good performance takes time and expends computational resources, it is often useful for a system to reason about its decision‐making capabilities. This article surveys a number of approaches for designing systems capable of managing their computational resources so as to maximize the expected utility derived from decision making. In particular, we consider methods that enable a system to trade some measure of decision‐making accuracy for speed in problem solving and planning. In some cases, the tradeoffs can be made at design time and compiled into a run‐time system. Alternatively, general knowledge for making the tradeoffs can be built into the system enabling it to make the necessary tradeoffs at run time. We explore the issues involved in making both sorts of tradeoffs and provide directions for further research.