Repository for the submission of NeurIPS Datasets and Benchmarks Track 2022.
The dataset in this repository includes three types of datasets for every game: random dataset, expert dataset and learning dataset. All dataset used in this work are avaiable at https://drive.google.com/drive/folders/1Y4hnkQ8hk2b81lbMaYWu26WTEF3L75FM?usp=sharing. The data entry in dataset is [current_game_state, player_id, legal_actions, action, next_game_state, next_legal_actions, next_player, reward, done, chance_node].
We also provide the code used to generate dataset in generate_dataset folder, therefore, the offline dataset of other games can be get using the code.
requirements.txt
train_bc_policy.py
file to get the behavior cloning policy by modifying the game and corresponding dataset in that filetrain_env_model.py
file to train the environment model by modifying the game and corresponding dataset in that file and then run OEF-CFR or OEF-PSRO algorithm to get the model-based policy based on the trained environment modelrun_mb_deep_cfr.py
file in the oef_cfr folder to run OEF-CFR algorithm on the trained environment modelrun_mb_psro.py
file in the oef_psro folder to run OEF-PSRO algorithm on the trained environment modelcombine_bc_mb.py
file to find proper weights for the behavior cloning policy and the model-based policy to get a best combination policy.