Multi-fidelity sampling for efficient simulation-based decision making in manufacturing management

Today’s manufacturers operate in highly dynamic and uncertain market environments. Process-level disturbances present further challenges. Consequently, it is of strategic importance for a manufacturing company to develop robust manufacturing capabilities that can quickly adapt to varying customer demands in the presence of external and internal uncertainty and stochasticity. Discrete-event simulations have been used by manufacturing managers to conduct “look-ahead” analysis and optimize resource allocation and production plan. However, simulations of complex manufacturing systems are time-consuming. Therefore, there is a great need for a highly efficient procedure to allocate a limited number of simulations to improve a system’s performance. In this article, we propose a multi-fidelity sampling algorithm that greatly increases the efficiency of simulation-based robust manufacturing management by utilizing ordinal estimates obtained from a low-fidelity, but fast, approximate model. We show that the multi-fidelity optimal sampling policy minimizes the expected optimality gap of the selected solution, and thus optimally uses a limited simulation budget. We derive an upper bound for the multi-fidelity sampling policy and compare it with other sampling policies to illustrate the efficiency improvement. We demonstrate its computational efficiency improvement and validate the convergence results derived using both benchmark test functions and two robust manufacturing management case studies.