This repository is the official implementation of the following paper accepted by the Conference on Uncertainty in Artificial Intelligence (UAI) 2021:
Quoc Phong Nguyen*, Zhaoxuan Wu*, Bryan Kian Hsiang Low, Patrick Jaillet
Trusted-Maximizers Entropy Search for Efficient Bayesian Optimization
To install requirements:
pip install cmake
pip install -r requirements.txt
As an example, optimizing the synthetic Branin function, run
bash script_batch_branin.shTo optimize real-world neural architecture search for CIFAR-10, run
bash script_batch_cifar.shThe code to optimize the real-world optimization problem of synthesizing faces to fool the python face_recognition library is not included.
NOTE: Configurations in the .sh files can be changed to fit different purposes. Some other pre-defined functions can be found in the functions.py file.
In this code repository, we implement:
- TES_sp: the configurations are
criterion=='sftl'andmode=='sample'; - TES_ep: the configurations are
criterion=='ftl'andmode=='ep'. However, the approximation with EP by matching the moments occasionally encounters numerical issues. Alternatively, we could resort to using samples to compute the moments:criterion=='ftl'andmode=='empirical'instead.