Tools to Perform Sobol Sensitivity Analysis on a Mathematical Model using a Gaussian Process Emulator.
Avoids performing Sobol sensitivity analysis directly on the model in question, as models often are expensive to evaluate. Instead, a statistical surrogate model is created from samples of the original models inputs and outputs. The statistical surrogate model is created using a Gaussian process emulator. Sobol sensitivity analysis is then performed on the surrogate model to obtain the first order (S1) and total order (ST) sensitivity indices.
- Implementation of Sobol sensitivity analysis using a Gaussian process emulator:
sensitivity_analysis_sobol.py. - Methods of training and making predictions using a Gaussian Process emulator:
gaussian_process_emulator.py. - An example data file containing the inputs and ouputs of an agent based model:
parameters_output.csv. - Example of how to extract Sobol sensitivities:
example_sobol.py. - Example of how to train and test a Gaussian process emulator:
example_gaussian_process_emulator.py. - Method to convert data with an esoteric structure (See bottom of README) to CSV:
esoteric_data_to_CSV.py.
Examples of (1) how to extract Sobol indices of a model from data and of (2) how to train and evaluate a Gaussian Process Emulator are provided.
(1) Running example_sobol.py analyses the parameter sensitivities from parameters_output.csv, producing following plot.
(2) Running example_gaussian_process_emulator.py trains a model on parameters_output.csv it then plots the models test set performance.
While superfluous to most users, esoteric_data_to_CSV.py is included specifically to parse data with the following structure.
├───UQ_eera_2020-09-11_part1
│ ├───output_row_0
│ │ └───output_prediction_simu_...
│ ├───output_row_1
│ │ └───output_prediction_simu_...
│ ...
...
To extract data from this type of structure and generate a CSV file use:
from esoteric_data_to_CSV import create_CSV_from_esoteric_data
create_CSV_from_esoteric_data("NAME_OF_OUTPUT_FILE.csv")