EURO-Online login
- New to EURO? Create an account
- I forgot my username and/or my password.
- Help with cookies
(important for IE8 users)
2905. Beyond Efficient Global Optimization: an Algorithm Minimizing Expected Regret
Invited abstract in session MB-27: On Mathematical Optimization for Explainable and Fair Machine Learning, stream Mathematical Optimization for XAI.
Monday, 10:30-12:00Room: 047 (building: 208)
Authors (first author is the speaker)
| 1. | Zizhou Ouyang
|
| Business School, University of Edinburgh | |
| 2. | Belen Martin Barragan
|
| University of Edinburgh Business School, The University of Edinburgh | |
| 3. | Xuefei Lu
|
| Skema Business School |
Abstract
Efficient Global Optimization (EGO) has been widely applied to simulation optimization problems, where it uses the Expected Improvement (EI) criterion to navigate the search space for optimal solutions, traditionally relying on Kriging for surrogate modelling and uncertainty estimation. Some existing efforts have attempted to broaden EGO's applicability by substituting Kriging with various machine learning models. However, approximating emulation uncertainty is limited in the absence of Kriging. In this paper, we propose the Expected Regret (ER) criterion, which captures uncertainty through data-driven insights without relying on Gaussian assumptions for uncertainty estimation and enables the direct utilization of machine learning regression models for surrogate modelling. We then apply the proposed algorithm to various simulators of different dimensionalities. Comparative performance evaluations with the classical EGO method demonstrate that our approach achieves competitive results while offering greater flexibility in selecting machine learning models, indicating its potential for addressing high-dimensional optimization problems.
Keywords
- Simulation
- Optimization Modeling
- Reliability
Status: accepted
Back to the list of papers