A SUFFICIENT CONDITION FOR NEAR-OPTIMAL STOCHASTIC CONTROLS AND ITS APPLICATION TO MANUFACTURING SYSTEMS

This paper presents an asymptotic analysis of control models governed by stochastic ordinary differential equations. A sufficient condition of near-optimal controls is given based on Ekeland's principle. It is shown that, under some concavity assumptions, the epsilon-maximum condition in terms of the Hamiltonian implies the square-root epsilon-optimality. By applying this result to a general manufacturing system, the common practice of ''hierarchical controls'' employed in order to reduce the overall complexity of the system is justified on a rigorous basis. A near-optimal control for the operational level is constructed from a near-optimal control at the corporate level, and an asymptotic error bound is obtained. A stochastic extension of the classical HMMS model is treated as a specific example. The approach of this paper is different from those in the literature, and it allows us to handle some previously unsolved problems with nonlinear state equations as well as nonseparable cost functions.