irelia

under construction...

review site(Version 1, not alpha go zero)

It is a review site that recorded a match between learned AIs with dataset, and it is the first version that use Q-table learning(dynamic state list).

It is not an algorithm of alpha go zero.

http://115.68.23.80:81/web/review.html

Prerequisite

• pip install colorama
• ...

self-play and train with no dataset

python self_play_and_train.py --save_dir="your path to save your model and self-play dataset" --max_step=100 --max_episode=10000 --max_simulation=200 --episode_interval_to_train=10 --print_mcts_tree=False --print_mcts_search=False

train with dataset

download the dataset

• 38MB
• Download the dataset

mkdir dataset
cd dataset
wget http://img.hovits.com/korean-chess-records-dataset.txt

convert the dataset to the real dataset for training

python parse_dataset.py --dataset_dir="the path you downloaded"

train with dataset

python optimizer.py --dataset_dir="the path you converted" --save_dir="your path to save your model" --epoch=10 --num_model_layers=20 --batch_size=32

Play with trained AI with MCTS

python user_vs_trained_mcts.py --save_dir="the model dir you trained" --model_file_name="the model name you trained" --max_step=100 --max_episode=10000 --max_simulation=200 --print_mcts_tree=False --print_mcts_search=False

Trained AI's self-play with MCTS

python play_trained_mcts_vs_trained_mcts.py --save_dir="the model dir you trained" --model_file_name="the model name you trained" --max_step=100 --max_episode=10000 --max_simulation=200 --print_mcts_tree=False --print_mcts_search=False

Trained AI's self-play with no MCTS

python play_net_vs_net.py --save_dir="the model dir you trained" --model_file_name="the model name you trained" --max_step=100 --max_episode=10000 --max_simulation=200 --print_mcts_tree=False --print_mcts_search=False