KL divergence – based disagreement sampling for multi-fidelity Bayesian optimization

Computational costs have increased tremendously over time owing to the increase in complexity of design. The motivation to reduce costs has led to the adoption of multi-fidelity optimization techniques. Active Learning has gained prominence as a sampling method which allows for maximum optimization of the design problem while utilizing minimal available data, consequently decreasing the cost of computation. This thesis proposes a novel KL-Divergence sampling strategy which iteratively makes use of fidelities to sample through a design space to optimize according to a target objective. This method was evaluated using the Rastrigin function and then applied on a design space consisting of 4-digit NACA airfoil combinations. It is observed that the KL divergence method provided a better optimization result compared to the Least Confidence method while simultaneously reducing reliance on information from higher fidelities. The disagreement sampler iterated over the airfoil design space and provided a better result as compared to the Least Confidence method.

Name thesis:degree_name

M.S.

Level thesis:degree_level

Thesis

Discipline thesis:degree_discipline

Aerospace Engineering

Grantor

University of Illinois Urbana-Champaign

Year dc:date

2025

Author dc:creator

• Iyer, Abhishek

Contributors dc:contributor

• Smart, Jordan

• Active Learning
• Bayesian Optimization
• Multi-fidelity
• Airfoil

Statement dc:rights

• Copyright 2025 Abhishek Iyer

Language dc:language

en, eng

Handle dc:identifier

https://hdl.handle.net/2142/129346

OAI record oai:identifier

https://www.ideals.illinois.edu/oai-pmh?verb=GetRecord&metadataPrefix=etdms&identifier=oai%3Awww.ideals.illinois.edu%3A2142%2F129346