We develop and analyze an algorithm to maximize the throughput of a serial kanban-based
manufacturing system with arbitrary
arrival and service process distributions by adjusting the number of kanban allocated to each
production stage while maintaining the total work-in-process inventory at any desired level. The
optimality properties of the algorithm are proved under a necessary and su$cient `smoothness
conditiona. The algorithm is driven by throughput sensitivities which, in general, can only be
estimated along an observed sample path of the system. It is shown that the algorithm converges to
the optimal allocation in probability and, under additional mild conditions, almost surely as
well. Finally, it is shown that Finite Perturbation Analysis (FPA) techniques can be used to obtain
the sensitivity estimates in order to reduce the amount of simulation required in either on-line
or o!-line simulation-based optimization. © 1999 Elsevier Science Ltd. All rights reserved.