Transfer Safe Sequential Learning (TSSL)

This repo contains classes and scripts to actively learn Gaussian Processes under safety critical scenario. This is attach to paper Global Safe Sequential Learning via Efficient Knowledge Transfer (arXiv).

Please reach out Cen-You.Li@de.bosch.com for any further information.

Setup

We use conda to manage our environment (see here for conda environment management).

Install python and the required packages:

conda env create --file environment.yml
conda activate tssl
pip install -e .

Our test files cover all important functions/classes of the repo. Run pytest to see if the environment is set up correctly:

conda activate tssl
pytest

If you see the following error message

OMP: Error #15: Initializing libiomp5md.dll, but found libiomp5md.dll already initialized.

try to remove path\to\user\.conda\envs\tssl\Lib\site-packages\torch\lib\libiomp5md.dll

Paper experiments

In the following, we describe how to reproduce the experiments in our paper.

Simulated datasets

We first generate the synthetic data, and then run experiments on the generated data.

Generate datasets:

python ./tssl/experiments/simulator/data_generator_safe_mogp.py --folder YOUR_DATA_PATH --dim 1
python ./tssl/experiments/simulator/data_generator_safe_mogp.py --folder YOUR_DATA_PATH --dim 2
python ./tssl/experiments/simulator/data_generator_safe_branin.py --folder YOUR_DATA_PATH
python ./tssl/experiments/simulator/data_generator_safe_hartmann3.py --folder YOUR_DATA_PATH
python ./tssl/experiments/simulator/data_generator_safe_multi_sources_branin.py --folder YOUR_DATA_PATH --num_sources 5

Now GP1D, GP2D, Branin, Hartmann3 and multi sources Branin data are generated.

With multi sources Branin, specify the max number of source tasks you want with --num_sources.

Run SAL

python ./tssl/experiments/safe_learning/main_safe_al_from_simulator.py \
 --experiment_data_dir YOUR_DATA_PATH \
 --experiment_output_dir YOUR_RESULT_PATH \
 --experiment_idx 1 \
 --simulator_config DESCRIBED_BELOW \
 --acquisition_function_config BasicSafePredEntropyAllConfig --label_safeland True --safe_lower 0 --safe_upper inf --n_data_initial 10 --n_steps 50 --n_data_test 200 --query_noisy True

--experiment_idx is the seed, we use [1-5]. --simulator_config specifies the dataset.

For GP1D data: use --simulator_config SingleTaskMOGP1Dz[0-19]Config.

For GP2D data: use --simulator_config SingleTaskMOGP2Dz[0-19]Config.

For Branin data: use --simulator_config SingleTaskBranin[0-4]Config.

For Hartmann3 data: use --simulator_config SingleTaskHartmann3_[0-4]Config.

Run EffTrans

python ./tssl/experiments/safe_learning/main_safe_transfer_al_from_simulator.py \
 --experiment_data_dir YOUR_DATA_PATH \
 --experiment_output_dir YOUR_RESULT_PATH \
 --experiment_idx 1 \
 --simulator_config DESCRIBED_BELOW \
 --kernel_config BasicMIAdditiveConfig --model_config BasicTransferGPModelConfig --acquisition_function_config BasicSafePredEntropyAllConfig --label_safeland True --safe_lower_source 0 --safe_upper_source inf --safe_lower 0 --safe_upper inf --n_data_s 100 --n_data_initial 10 --n_steps 50 --n_data_test 200 --query_noisy True

--experiment_idx is the seed, we use [1-5]. --simulator_config specifies the dataset.

For GP1D data: use --simulator_config TransferTaskMOGP1Dz[0-19]Config --n_data_s 100 --n_data_initial 10 --n_steps 50.

For GP2D data: use --simulator_config TransferTaskMOGP2Dz[0-19]Config --n_data_s 250 --n_data_initial 20 --n_steps 100 --n_data_test 500.

For Branin data: use --simulator_config TransferTaskBranin[0-4]Config --n_data_s 100 --n_data_initial 20 --n_steps 100 --n_data_test 500.

For Hartmann3 data: use --simulator_config TransferTaskHartmann3_[0-4]Config --n_data_s 100 --n_data_initial 20 --n_steps 100 --n_data_test 500 --label_safeland False.

For multi-sources-Branin, run (we use --dim_s= 3 or 4, which is the number of source tasks)

python ./tssl/experiments/safe_learning/main_multi_sources_safe_transfer_al_from_simulator.py \
 --experiment_data_dir YOUR_DATA_PATH \
 --experiment_output_dir YOUR_RESULT_PATH \
 --experiment_idx [1-5] \
 --simulator_config TransferTaskMultiSourcesBranin[0-4]Config \
 --kernel_config BasicMIAdditiveConfig --model_config BasicTransferGPModelConfig --acquisition_function_config BasicSafePredEntropyAllConfig --label_safeland True --safe_lower_source 0 --safe_upper_source inf --safe_lower 0 --safe_upper inf --n_data_s_per_dim 20 --dim_s 3 --n_data_initial 20 --n_steps 100 --n_data_test 500 --query_noisy True

Run FullTransHGP:

replace --model_config BasicTransferGPModelConfig by --model_config BasicSOMOGPModelConfig, the remaining arguments are the same as EffTrans

Run FullTransLMC:

replace --kernel_config BasicMIAdditiveConfig --model_config BasicTransferGPModelConfig by --kernel_config BasicCoregionalizationPLConfig --model_config BasicSOMOGPModelConfig, the remaining arguments are the same as EffTrans

Run Rothfuss et al. 2022

replace --kernel_config BasicMIAdditiveConfig --model_config BasicTransferGPModelConfig by --kernel_config BasicFPACOHKernelConfig --model_config BasicMetaGPModelConfig, the remaining arguments are the same as EffTrans

engine datasets

The datasets are published on github (link).

First move all the excel files into path YOUR_DATA_PATH/engine

Run SAL

python ./tssl/experiments/safe_learning/main_safe_al_from_simulator.py \
 --experiment_data_dir YOUR_DATA_PATH \
 --experiment_output_dir YOUR_RESULT_PATH \
 --experiment_idx 1 \
 --simulator_config SingleTaskEngineInterpolated_PI0s_Config --acquisition_function_config BasicSafePredEntropyAllConfig --safe_lower -100 --safe_upper 1.0 --n_data_initial 20 --n_steps 100 --n_data_test 1000  --label_safeland True --query_noisy False

--experiment_idx is the seed, we use [1-5]. --simulator_config SingleTaskEngineInterpolated_TEx_Config can also be used (supplementary experiment).

Run EffTrans

python ./tssl/experiments/safe_learning/main_safe_transfer_al_from_simulator.py \
 --experiment_data_dir YOUR_DATA_PATH \
 --experiment_output_dir YOUR_RESULT_PATH \
 --experiment_idx 1 \
 --simulator_config TransferTaskEngineInterpolated_PI0s_Config --kernel_config BasicMIAdditiveConfig --model_config BasicTransferGPModelConfig --acquisition_function_config BasicSafePredEntropyAllConfig --safe_lower_source -100 --safe_upper_source 100 --safe_lower -100 --safe_upper 1.0 --n_data_s 500 --n_data_initial 20 --n_steps 100 --n_data_test 1000  --label_safeland True --query_noisy False

--experiment_idx is the seed, we use [1-5]. --simulator_config TransferTaskEngineInterpolated_TEx_Config can also be used (supplementary experiment).

Run FullTransHGP

python ./tssl/experiments/safe_learning/main_safe_transfer_al_from_simulator.py \
 --experiment_data_dir YOUR_DATA_PATH \
 --experiment_output_dir YOUR_RESULT_PATH \
 --experiment_idx 1 \
 --simulator_config TransferTaskEngineInterpolated_PI0s_Config --kernel_config BasicMIAdditiveConfig --model_config BasicSOMOGPModelConfig --acquisition_function_config BasicSafePredEntropyAllConfig --safe_lower_source -100 --safe_upper_source 100 --safe_lower -100 --safe_upper 1.0 --n_data_s 500 --n_data_initial 20 --n_steps 100 --n_data_test 1000  --label_safeland True --query_noisy False

--experiment_idx is the seed, we use [1-5]. --simulator_config TransferTaskEngineInterpolated_TEx_Config can also be used (supplementary experiment).

Run FullTransLMC

python ./tssl/experiments/safe_learning/main_safe_transfer_al_from_simulator.py \
 --experiment_data_dir YOUR_DATA_PATH \
 --experiment_output_dir YOUR_RESULT_PATH \
 --experiment_idx 1 \
 --simulator_config TransferTaskEngineInterpolated_PI0s_Config --kernel_config BasicCoregionalizationPLConfig --model_config BasicSOMOGPModelConfig --acquisition_function_config BasicSafePredEntropyAllConfig --safe_lower_source -100 --safe_upper_source 100 --safe_lower -100 --safe_upper 1.0 --n_data_s 500 --n_data_initial 20 --n_steps 100 --n_data_test 1000  --label_safeland True --query_noisy False

--experiment_idx is the seed, we use [1-5]. --simulator_config TransferTaskEngineInterpolated_TEx_Config can also be used (supplementary experiment).

Run Rothfuss et al. 2022

python ./tssl/experiments/safe_learning/main_safe_transfer_al_from_simulator.py \
 --experiment_data_dir YOUR_DATA_PATH \
 --experiment_output_dir YOUR_RESULT_PATH \
 --experiment_idx 1 \
 --simulator_config TransferTaskEngineInterpolated_PI0s_Config --kernel_config BasicFPACOHKernelConfig --model_config BasicMetaGPModelConfig --acquisition_function_config BasicSafePredEntropyAllConfig --safe_lower_source -100 --safe_upper_source 100 --safe_lower -100 --safe_upper 1.0 --n_data_s 500 --n_data_initial 20 --n_steps 100 --n_data_test 1000 --label_safeland True --query_noisy False

--experiment_idx is the seed, we use [1-5]. --simulator_config TransferTaskEngineInterpolated_TEx_Config can also be used (supplementary experiment).