Deep (Predictive) Discounted Counterfactual Regret Minimization

Deep (Predictive) Discounted Counterfactual Regret Minimization
Hang Xu, Kai Li^#, Haobo Fu, Qiang Fu, Junliang Xing, Jian Cheng
AAAI 2026 (Oral)

Install DeepPDCFR

Install miniconda3 from the official website and run the following script:

bash scripts/install.sh

Download the file flop_matrix.npy from this url, and put it into the folder matrix.

Train DeepPDCFR

Run the following script to assess the performance of model-free neural algorithms on testing games. The results are saved in the folder logs.

conda activate DeepPDCFR
python scripts/run.py with configs/{algo_name}.yaml game_name={game_name} seed={seed} --force

algo_name is the algorithm name chosen from NFSP, QPG, RPG, OSDeepCFR, VRDeepDCFRPlus, VRDeepPDCFRPlus.

game_name is the testing game name chosen from KuhnPoker, LeducPoker, LiarsDice5, LiarsDice6, GoofSpielImp5, GoofSpielImp6, Battleship_22_3, Battleship_32_3, FHP

seed is the random seed chosen from 0, 1, 2, 3.