Robust system design with limited experimental data and an inexact simulation model

Computer simulations of physical systems are commonly used to improve system performance via optimizing the system design. We study the case when, in addition to the simulation, a limited amount of experimental data is collected from the real physical system. This article describes a method for selecting a conservative system design that is robust to uncertainty from the simulation parameters and simulation bias. The concept is that each potential system design is assigned a worst-case scenario in a data-driven feasible region. The conservative system design is then chosen as the best of the worst-cases. The method is shown to have good statistical properties. A case study is performed where a vehicle safety belt design is chosen to minimize the impact of vehicle crashes on a driver.