Instructions for running the code

To run POELA:

python src/run_pg.py --action_mask_type=nn_action_dist --threshold=0.6 --var_coeff=0.1

To run PO-mu/PO-CRM:

python src/run_pg.py --action_mask_type=step --threshold=0.01 --var_coeff=10.0

To run BCQ:

python src/run_ql.py --state_clipping=0 --threshold=0.01

To run PQL:

python src/run_ql.py --state_clipping=1 --threshold=0.01

Acknowledgements

The code is an adaptation of the BCQ official implementation.

This is code for the paper Offline Policy Optimization with Eligible Actions https://arxiv.org/abs/2207.00632 Yao Liu, Yannis Flet-Berliac, and Emma Brunskill Conference on Uncertainty in AI (UAI) 2022