Suzume-Jong with reinforcement learning
folder description
baseline_results/ save 1player optimal(greedy) play (or random play) results
baseline_run.py run 1player optimal(greedy) play in Suzume-jan, and save log
first_suzume.ipynb this file made hand2optimal, hand_idx_prob and value_iteration
hand2optimal/ npy or pickle files to get optimal discard tiles
mahjong_networks.py define custom network(mlp, conv1d, conv2d, resnet)
mahjong_utils.py suzume-jan utility function such as is_win, calc_point and so on
reward_predict.py make global reward predictor
suzume_env/ Suzume-jan envirionment
train_ppo.py train and test ppo
value_iteration/ save Q-values by using value iteration
OpenAI baselines, tensorflow==1.14, keras
- test baseline player and save the result to
baseline_resultsdir
python baseline_run.py [-p NUM_PLAYER] [-t TOTAL_TIMESTEPS] [-og OPPONENT_PLAYER_GAMMA] [-pg PLAYER_GAMMA] [-rs REWARD_SCALE] [-pr]
python baseline_run.py -p 2 -t 300000 -og 09 -pg random -rs 20 -ps
- show the result
python baseline_run.py [-l PICKLE_NAME]
python baseline_run.py -l baseline_results/2_09_09_3000000.pickle
- train ppo algorithm
python train_ppo.py [--env ENV_NAME] [--save_path PICKLE_PATH] [-t NUM_TIMESTEPS]
python train_ppo.py --env Suzume2-v0 --save_path save_models/default.pkl -t 30000
- show result (when --play, set num_timesteps to 0 and num_env to 1)
python train_ppo.py --play [--load_path MODEL_PATH]
python train_ppo.py --play --load_path save_models/ppo_2_5_1_99.pkl --pickle_name 09
for more details, run python train_ppo.py -h
- make global reward predictor (save results to
save_rew_preddir)
python reward_predict.py [-p NUM_PLAYERS] [-pg PLAYER_GAMMA] [-og OPPOPNENT_GAMMA] [-m MODEL_NAME] [--play]
python reward_predict.py -m dense
This project is licensed under the Apache License, Version 2.0. See LICENSE for the full license text.