Good-Action Identification

Code for ICML 2021 paper Lenient Regret and Good-Action Identification in Gaussian Process Bandits

Copyright by the authors: Xu Cai, Selwyn Gomes and Jonathan Scarlett

Dependencies

• Python 3
• NumPy
• SciPy
• Scikit-Learn
• Matplotlib
• GPy
• PyTorch (for quasi-random sequences)
• pybox2d & pygame (for robot pushing)
• xgboost (for XGBoost)

The noisy/noiseless experiment on synthetic/real-world functions

Input arguments for main.py:

• function: Specify the function name; See good_action/utils.py for details
• noisy: Noisy or noiseless observation
• epsilon: The good-action threshold; Float value

Output:

• log file: Running logs
• query histories: .npy file saving queried points and values

For example:

• Testing on the noiseless 3D robot pushing function

python main.py robot3 noiseless 4.5

Visualization: Run plot.ipynb

The lenient regret experiment on synthetic GP function

Input arguments for lenient.py:

• epsilon: The good-action threshold; Float value; Default=0.9

Output:

• lenient and standard regrets: .npy file

For example:

python lenient.py 0.9

Visualization: Run plot_lenient.ipynb