Neuro-dynamic programming is a class of powerful techniques for approximatingthe solution to dynamic programming equations. In their most computationallyattractive formulations, these techniques provide the approximate solution onlywithin a prescribed finite-dimensional function class. Thus, the question thatalways arises is how should the function class be chosen? The goal of thispaper is to propose an approach using the solutions to associated fluid anddiffusion approximations. In order to illustrate this approach, the paperfocuses on an application to dynamic speed scaling for power management incomputer processors.
展开▼
