TY - GEN
T1 - A Framework for Multi-fidelity Modeling in Global Optimization Approaches
AU - Zabinsky, Zelda B.
AU - Pedrielli, Giulia
AU - Huang, Hao
N1 - Funding Information:
in part by the National Science Foundation, Grant
Funding Information:
This work has been supported in part by the National Science Foundation, Grant CMMI-1632793.
Publisher Copyright:
© Springer Nature Switzerland AG 2019.
PY - 2019
Y1 - 2019
N2 - Optimization of complex systems often involves running a detailed simulation model that requires large computational time per function evaluation. Many methods have been researched to use a few detailed, high-fidelity, function evaluations to construct a low-fidelity model, or surrogate, including Kriging, Gaussian processes, response surface approximation, and meta-modeling. We present a framework for global optimization of a high-fidelity model that takes advantage of low-fidelity models by iteratively evaluating the low-fidelity model and providing a mechanism to decide when and where to evaluate the high-fidelity model. This is achieved by sequentially refining the prediction of the computationally expensive high-fidelity model based on observed values in both high- and low-fidelity. The proposed multi-fidelity algorithm combines Probabilistic Branch and Bound, that uses a partitioning scheme to estimate subregions with near-optimal performance, with Gaussian processes, that provide predictive capability for the high-fidelity function. The output of the multi-fidelity algorithm is a set of subregions that approximates a target level set of best solutions in the feasible region. We present the algorithm for the first time and an analysis that characterizes the finite-time performance in terms of incorrect elimination of subregions of the solution space.
AB - Optimization of complex systems often involves running a detailed simulation model that requires large computational time per function evaluation. Many methods have been researched to use a few detailed, high-fidelity, function evaluations to construct a low-fidelity model, or surrogate, including Kriging, Gaussian processes, response surface approximation, and meta-modeling. We present a framework for global optimization of a high-fidelity model that takes advantage of low-fidelity models by iteratively evaluating the low-fidelity model and providing a mechanism to decide when and where to evaluate the high-fidelity model. This is achieved by sequentially refining the prediction of the computationally expensive high-fidelity model based on observed values in both high- and low-fidelity. The proposed multi-fidelity algorithm combines Probabilistic Branch and Bound, that uses a partitioning scheme to estimate subregions with near-optimal performance, with Gaussian processes, that provide predictive capability for the high-fidelity function. The output of the multi-fidelity algorithm is a set of subregions that approximates a target level set of best solutions in the feasible region. We present the algorithm for the first time and an analysis that characterizes the finite-time performance in terms of incorrect elimination of subregions of the solution space.
KW - Gaussian processes
KW - Global optimization
KW - Meta-models
KW - Multi-fidelity models
KW - Probabilistic Branch and Bound
UR - http://www.scopus.com/inward/record.url?scp=85078465300&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85078465300&partnerID=8YFLogxK
U2 - 10.1007/978-3-030-37599-7_28
DO - 10.1007/978-3-030-37599-7_28
M3 - Conference contribution
AN - SCOPUS:85078465300
SN - 9783030375980
T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
SP - 335
EP - 346
BT - Machine Learning, Optimization, and Data Science - 5th International Conference, LOD 2019, Proceedings
A2 - Nicosia, Giuseppe
A2 - Pardalos, Panos
A2 - Umeton, Renato
A2 - Giuffrida, Giovanni
A2 - Sciacca, Vincenzo
PB - Springer
T2 - 5th International Conference on Machine Learning, Optimization, and Data Science, LOD 2019
Y2 - 10 September 2019 through 13 September 2019
ER -