A sparse multi-fidelity surrogate-based optimization method with computational awareness

Haizhou Yang

doi:10.1007/s00366-022-01766-8

Return

A sparse multi-fidelity surrogate-based optimization method with computational awareness

2022-12-01

PRE

Haizhou Yang

王

王毅 (Yi Wang) *

DOI：10.1007/s00366-022-01766-8

Original

Original request for help

Save

Abstract

En 中文

CoKriging is a popular surrogate modeling approach to approximate the input-output relationship using multi-fidelity data from different sources. However, it suffers from the big data issue due to its cubic time complexity and square memory complexity with the data size. This becomes even exacerbated for iterative design optimization. To overcome this limitation, this paper presents a new data sparsification method for multi-fidelity surrogate-based optimization (MFSBO). It includes two key components: reduced design space and data filtering (RDS & DF), which alleviate the surrogate modeling complexity and time for improved efficiency while balancing between exploration and exploitation during optimization. RDS & DF is also combined with an expected improvement reduction (EIR)-based infill technique (Yang et al. in Struct Multidiscip Optim, 2022. https://doi.org/10.1007/s00158-022-03240-x), enabling both parsimony and computational awareness for MFSBO. Two case studies are conducted to examine the proposed method. Results demonstrate that a significant reduction in the modeling and thus optimization time (69.49% and 73.92%) is achieved while retaining the design accuracy.

Keywords:

Big data

Gaussian process

CoKriging

Reduced design space

Microfluidic concentration gradient generator