Suboptimal Design for Large Scale, Multimodule Systems

We present a general procedure for determining computationally tractable suboptimal designs, associated upper and lower bounds on the optimal expected cost function, and a procedure for determining a suboptimal design which improves these bounds for an important class of large scale, stochastic decision processes. This class, called the multimodule Markov decision process, is distinguished by the characteristic that each element of the vector state process has dynamics that operate independently of the other elements. The general suboptimal design procedure is based on the solution of a subglobal optimization problem for each module, where for each subglobal problem it is assumed that only restricted state information from the composite system is available to its associated module-decision maker. The computational implications of this suboptimal design approach are analyzed. A multi-component replacement example illustrates this suboptimal design procedure, its associated bounds, and a suboptimal procedure which improves these bounds.