Trusted-Maximizers Entropy Search for Efficient Bayesian Optimization

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

Requirements

To install requirements:

pip install cmake
pip install -r requirements.txt

Running the scripts

As an example, optimizing the synthetic Branin function, run

bash script_batch_branin.sh

To optimize real-world neural architecture search for CIFAR-10, run

bash script_batch_cifar.sh

The 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.

Remarks

In this code repository, we implement:

• TES_sp: the configurations are criterion=='sftl' and mode=='sample';
• TES_ep: the configurations are criterion=='ftl' and mode=='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' and mode=='empirical' instead.