MULTI-FIDELITY OPTIMIZATION WITH GAUSSIAN REGRESSION ON ORDINAL TRANSFORMATION SPACE

Abstract

In the areas of Discrete Simulation Optimization, optimization over rank values instead of actual values has received increasing attention in recent years. The concept of Ordinal Transformation (OT) has been proposed with the use of multi-fidelity models, capable of transforming the design space of the simulation optimization problem in question into a single dimension space. The dimension reduction is able to smooth out potentially irregular design space, allowing meta-model approaches of sampling. On the other hand, Gaussian Process Regression (GPR) has been widely researched and adopted in machine learning as well as simulation, given its powerful performance in capturing the general structure of the system. In this work we will discuss the application of combining GPR with OT, the assumptions for this strategy to work, and the implications to the performance. A sampling strategy based on the two will be derived and expanded to more than one low fidelity cases.