Thesis
Multi-fidelity blackbox optimization of continuous spaces: An empirical study
Washington State University
Master of Science (MS), Washington State University
05/2019
DOI:
https://doi.org/10.7273/000004014
Handle:
https://hdl.handle.net/2376/125196
Abstract
Many real-life problems require optimizing functions with expensive evaluations. Bayesian Optimization (BO) and Optimistic Optimization (OO) are two broad families of algorithms that try to find the global optima of a function with the goal of minimizing the number of function evaluations. A large body of existing work deals with the single-fidelity setting, where function evaluations are very expensive but accurate. However, in many applications (e.g., hyper-parameter tuning of machine learning methods), we have access to multiple-fidelity functions that vary in their cost and accuracy of evaluation. In this dissertation, we perform a rigorous empirical study of existing multi-fidelity optimization algorithms from both BO and OO families towards the goal of identifying their strengths and weaknesses. We also study hybrid algorithms by combining the best attributes of both BO and OO algorithms to discover the global optima of a blackbox function with minimal cost. Our experiments on multiple benchmark functions and hyper-parameter tuning problems reveal several interesting insights regarding the behavior of different multi-fidelity algorithms, and show that hybrid algorithms generally perform better than BO and OO methods. Hybrid algorithms employ BO principles to to guide the search towards most promising regions to uncover the optima and employs OO principles to safely prune large parts of the search spaces, thereby uncovering better solutions using less resources for function evaluations than BO and OO based approaches. Our experiments reveal that that the best performing algorithms employ more lower-fidelity function evaluations to prune large parts of the search space.
Metrics
8 File views/ downloads
19 Record Views
Details
- Title
- Multi-fidelity blackbox optimization of continuous spaces
- Creators
- Mohammad Omar Faruk
- Contributors
- Venkata Janardhan Rao Doppa (Advisor) - Washington State University, School of Electrical Engineering and Computer Science
- Awarding Institution
- Washington State University
- Academic Unit
- School of Electrical Engineering and Computer Science
- Theses and Dissertations
- Master of Science (MS), Washington State University
- Publisher
- Washington State University
- Identifiers
- 99900890796601842
- Language
- English
- Resource Type
- Thesis