We introduce SCOUT-Nd (Stochastic Constrained Optimization for N Dimensions) and MF-SCOUT-Nd (Multi-Fidelity Stochastic Constrained Optimization for N Dimensions) for constrained stochastic optimization involving stochastic black-box physics-based simulators with high-dimensional parametric dependency. The proposed algorithm consists of the following major elements:
- A non-intrusive method to estimate gradients of black-box physical simulators, with an ability to account for stochasticity in the objective and handle constraints using penalty methods.
- Strategies to reduce the variance of the gradient estimator.
- Ability to handle non-convexity.
- Better optimum and well-behaved convergence properties using natural gradients.
- Multi-fidelity strategies and adaptive selection of the number of samples for gradient estimation to provide trade-off between computational cost and accuracy.
The scripts to reproduce the studies in the paper will be made available upon publication.
Benchmark codes here
- Windfarm layout optimization here
- Pipe shape optimization code, student thesis
If you use this code, please cite our paper:
@article{agrawal:hal-04659802,
TITLE = {{Stochastic Black-Box Optimization using Multi-Fidelity Score Function Estimator}},
AUTHOR = {Agrawal, Atul and Koutsourelakis, Phaedon-Stelios and Ravi, Kislaya and Bungartz, Hans-Joachim},
URL = {https://hal.science/hal-04659802},
NOTE = {working paper or preprint},
YEAR = {2024},
MONTH = Jul,
PDF = {https://hal.science/hal-04659802/file/scout_nd_preprint_neurips_formal_ml_scitech.pdf},
HAL_ID = {hal-04659802},
HAL_VERSION = {v1},
}