Abstract
A control system operating in a complex environment will encounter a variety of different situations, with varying amounts of time available to respond to critical events. Ideally, such a control system will do the best possible with the time available. In other words, its responses should approximate those that would result from having unlimited time for computation, where the degree of the approximation depends on the amount of time it actually has. There exist approximation algorithms for a wide variety of problems. Unfortunately, the solution to any reasonably complex control problem will require solving several computationally intensive problems. Algorithms for successive approximation are a subclass of a class of anytime algorithms, algorithms that return answers for any amount of computation time, where the answeres improve as more time is allotted. In this paper, we describe an architecture for allocating computation time to a set of anytime algorithms, based on expectations regarding the value of the answers they return. The architecture we describe is quite general, producing optimal chedules for a set of algorithms under widely varying conditions.