Active Learning for Derivative-Based Global Sensitivity Analysis with Gaussian Processes

This repository includes all the methods proposed and discussed in the paper Active Learning for Derivative-Based Global Sensitivity Analysis with Gaussian Processes.

Below are steps to run the code:

Set up enviroment

conda create -y --name gsa python=3.8
conda run -n gsa pip install mpmath 
conda run -n gsa pip install ax-platform 
conda activate gsa

Launch one run with a single seed:

mkdir results/function_name
python run.py function_name method_name num_iter num_init seed

Launch a multi-runs experiment with several seeds:

bash experiments.sh function_name method_name num_iter num_init repeats

• List of implemented methods names:

  • Sobol: Quasi random Monte Carlo Sequence ($QR$).

  • MaxVariance: GP variance ($fVAR$).

  • InformationGain: GP information gain ($fIG$).

  • DerivVarianceTrace: Variance of the derivative GP ($DV$).

  • DerivVarianceTraceReduction: Variance reduction of the derivative GP ($DV_r$).

  • DerivSumInformationGain: Information gain of the derivative GP ($DIG$).

  • DerivAbsVarianceTrace: Variance of the absolute of the derivative GP ($DAbV$).

  • DerivAbsVarianceTraceReduction: Variance reduction of the absolute of the derivative GP ($DAbV_r$).

  • DerivAbsSumInformationGain{i}: Information gain of the absolute of the derivative GP using the $i$ th approximation ($DAbIG_i$).

  • DerivSquareVarianceTrace: Variance of the square of the derivative GP ($DSqV$).

  • DerivSquareVarianceTraceReduction: Variance reduction of the square of the derivative GP ($DSqV_r$).

  • DerivSquareSumInformationGain{i}: Information gain of the square of the derivative GP using the $i$ th approximation ($DSqIG_i$).

  • GlobalDerivVarianceTraceReduction: Global (integrated) variance reduction of the derivative GP ($GDV_r$).

  • GlobalDerivSumInformationGain: Global (integrated) information gain of the derivative GP ($GDIG$).

  • GlobalDerivAbsVarianceTraceReduction: Global (integrated) variance reduction of the absolute of the derivative GP ($GDAbV_r$).

  • GlobalDerivAbsSumInformationGain{i}: Global (integrated) information gain of the square of the derivative GP using the $i$ th approximation ($GDAbIG_i$).

  • GlobalDerivSquareVarianceTraceReduction: Global (integrated) variance reduction of the square of the derivative GP ($GDSqV_r$).

  • GlobalDerivSquareSumInformationGain{i}: Global (integrated) information gain of the square of the derivative GP using the $i$ th approximation ($GDSqIG_i$).

• Test functions: The list of implemented function names can be found in utils/util.py and their corresponding true DGSMs are in utils/dgsm_values.py

• Note: This repository also includes an interface to call the acquisition functions proposed Wycoff et al 2021. The methods names for these acquisition functions are activegp:Cvar, activegp:Cvar2 and activegp:Ctr. For these approaches to work, the requirements in corresponding R package have to be installed.

Citation

If you use this code, please consider citing our paper:

@article{Belakaria_Letham_2024, 
  title={Active Learning for Derivative-Based Global Sensitivity Analysis with Gaussian Processes}, 
  url={},
  journal={Advances in neural information processing systems}, 
  author={Belakaria, Syrine and Letham, Benjamin and  Doppa, Janardhan Rao and Engelhardt, Barbara and Ermon, Stefano and Bakshy, Eytan}, 
  year={2024}, 
  month={June}}