An effective learning procedure for multi-fidelity simulation optimization with ordinal transformation

Simulation models of different fidelity levels are often available for the same complex system. High-fidelity models generate accurate measurements of the performance of a system design but can only be simulated for a very limited number of designs due to its prohibitively expensive computation cost. In contrast, low-fidelity models produce approximate estimates of the objective function but are lightweight and can evaluate a large number of designs in a short amount of time. Ordinal transformation is a novel framework that combines the merits of high- and low-fidelity simulation models to perform efficient optimization. In this paper, we propose an effective learning procedure that improves the prediction accuracy of low-fidelity models. Numerical experiment demonstrates the promising performance of learning within the ordinal transformation framework.