This work presents a preliminary attempt to address a fundamental question for decision-making: Can we develop a single algorithm to tackle ALL categories of decision-making problems?
As this code repository uses the package OpenSpiel, an environment with the Python version that is compatible with OpenSpiel is preferred, e.g., Python >= 3.7.
All the game environments are built on top of OpenSpiel, which can be installed using the following command:
python3 -m pip install open_spielTo run the experiments, just execute the following command in the command line:
sh train.shThe command for execution is in the script file train.sh. All other hyperparameters and their default values can be found in the file train.py. One can easily run different methods by changing the corresponding hyperparameter values.
game="kuhn_poker(players=3)" # specify the game
solver="cmd" # use MD-type or cfr-type algorithm
comb="gmd" # use KL/EU (closed-form solution) or GMD (numerical method)
alpha_sc="mtctl" # use meta-controller to adjust alpha or fix the alpha
alpha_optim="drs" # the meta-controller used to adjust alpha (other choices: dgld, rs, ...)
alpha_optim_obj="negap" # the evaluation metric of interest (other choices: sw, ccegap)
phi=3 # the choice of Bregman divergence/convex function
seed=1 # random seed
python train.py --game ${game} --solver ${solver} --comb ${comb} --alpha_sc ${alpha_sc} \
--alpha_optim ${alpha_optim} --alpha_optim_obj ${alpha_optim_obj} \
--phi ${phi} --seed ${seed}
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent