Dynamic Scheduling of Progressive Processing Plans

Progressive processing plans allow systems to tradeoff computational resources against the quality of service by specifying alternative ways in which to accomplish each step. When the structure of a plan is known in advance, it can be optimally scheduled by solving a corresponding Markov decision process. This paper extends this approach to dynamic scheduling of plans that can be constantly modified. We show how to construct an optimal meta-level controller for a single task and how to extend the solution to the case of multiple and dynamic tasks using the notion of an opportunity cost. Several fast approximation schemes for the opportunity cost are evaluated. The results provide an effective framework for managing computational resources in highly dynamic environments.