setup the conda environment using the file env.yml

Navigate to our_method/visualization_scipts/ to use existing trained models to generate trajectories:

• simulation_latest_for_gt_comparison simulates the trajectories for the 4-stage game
• simulation_latest.py simulates the unconstrained case
• simulation_latest_primal_dual.py simulates the unconstrained case with both primal and dual policies
• simulation_latest_cons.py simulates the constrained case
• simulation_latest_cons_primal_dual.py simulates the constrained case with both primal and dual policies

Navigate to our_method/ to train the value network for different cases -- unconstrained, contrained and their dual versions and the 3d case.

• run ./train_our_method.sh to train the primal unconstrained case
• run ./train_our_method_for_cfr.sh to train the comparison case (against DeepCFR)
• run ./train_our_method_dual.sh to train the dual unconstrained case
• run ./train_our_method_cons.sh to train the primal constrained case
• run ./train_our_method_cons_dual.sh to train the dual constrained case
• run ./train_our_method_3d.sh to train the primal high dimensional case

To train deep cfr policy networks, run run_cfr_3.py for $|A|=9$, and run_cfr for $|A|=16$.

To compare our method with cfr, run the notebook our_method/hexner_last_step-stopping.ipynb

To generate trajectories using deepcfr, run the notebook DeepCFR_Trajectory.ipynb