Chinese Chess reinforcement learning by AlphaZero methods.

This project is based on these main resources:

1. DeepMind's Oct 19th publication: Mastering the Game of Go without Human Knowledge.
2. The great Reversi/Chess/Chinese chess development of the DeepMind ideas that @mokemokechicken/@Akababa/@TDteach did in their repo: https://github.com/mokemokechicken/reversi-alpha-zero, https://github.com/Akababa/Chess-Zero, https://github.com/TDteach/AlphaZero_ChineseChess
3. A Chinese chess engine with gui: https://github.com/mm12432/MyChess

In order to build a strong chinese chess AI following the same type of techniques as AlphaZero, we need to do this with a distributed project, as it requires a huge amount of computations.

If you want to join us to build the best chinese chess AI in the world:

• For instructions, see wiki
• For live status, see https://cczero.org

• Python 3.6.3
• tensorflow-gpu: 1.3.0
• Keras: 2.0.8

This AlphaZero implementation consists of two workers: self and opt.

• self is Self-Play to generate training data by self-play using BestModel.
• opt is Trainer to train model, and generate new models.

For the sake of faster training, another two workers are involved:

• sl is Supervised learning to train data crawled from the Internet.
• eval is Evaluator to evaluate the NextGenerationModel with the current BestModel.

Requirement: pygame

python cchess_alphazero/run.py play

Screenshots

You can choose different board/piece styles and sides, see play with human.

pip install -r requirements.txt

If you want to use CPU only, replace tensorflow-gpu with tensorflow in requirements.txt.

Make sure Keras is using Tensorflow and you have Python 3.6.3+.

PlayDataConfig

• nb_game_in_file, max_file_num: The max game number of training data is nb_game_in_file * max_file_num.

PlayConfig, PlayWithHumanConfig

• simulation_num_per_move : MCTS number per move.
• c_puct: balance parameter of value network and policy network in MCTS.
• search_threads: balance parameter of speed and accuracy in MCTS.
• dirichlet_alpha: random parameter in self-play.

python cchess_alphazero/run.py self

When executed, Self-Play will start using BestModel. If the BestModel does not exist, new random model will be created and become BestModel.

options

• --new: create new BestModel
• --type mini: use mini config, (see cchess_alphazero/configs/mini.py)
• --gpu '1': specify which gpu to use
• --ucci: whether to play with ucci engine (rather than self play, see cchess_alphazero/worker/play_with_ucci_engine.py)
• --distributed: run self play in distributed mode which means it will upload the play data to the remote server and download latest model from it

python cchess_alphazero/run.py opt

When executed, Training will start. The current BestModel will be loaded. Trained model will be saved every epoch as new BestModel.

options

• --type mini: use mini config, (see cchess_alphazero/configs/mini.py)
• --total-step TOTAL_STEP: specify total step(mini-batch) numbers. The total step affects learning rate of training.
• --gpu '1': specify which gpu to use

View training log in Tensorboard

tensorboard --logdir logs/

And access http://<The Machine IP>:6006/.

python cchess_alphazero/run.py play

When executed, the BestModel will be loaded to play against human.

options

• --ai-move-first: if set this option, AI will move first, otherwise human move first.
• --type mini: use mini config, (see cchess_alphazero/configs/mini.py)
• --gpu '1': specify which gpu to use
• --piece-style WOOD: choose a piece style, default is WOOD
• --bg-style CANVAS: choose a board style, default is CANVAS
• --cli: if set this flag, play with AI in a cli environment rather than gui

Note: Before you start, you need to download/find a font file (.ttc) and rename it as PingFang.ttc, then put it into cchess_alphazero/play_games. I have removed the font file from this repo because it's too big, but you can download it from here.

python cchess_alphazero/run.py eval

When executed, evaluate the NextGenerationModel with the current BestModel. If the NextGenerationModel does not exist, worker will wait until it exists and check every 5 minutes.

options

• --type mini: use mini config, (see cchess_alphazero/configs/mini.py)
• --gpu '1': specify which gpu to use

python cchess_alphazero/run.py sl

When executed, Training will start. The current SLBestModel will be loaded. Tranined model will be saved every epoch as new SLBestModel.

About the data

I have two data sources, one is downloaded from https://wx.jcloud.com/market/packet/10479 ; the other is crawled from http://game.onegreen.net/chess/Index.html (with option --onegreen).

options

• --type mini: use mini config, (see cchess_alphazero/configs/mini.py)
• --gpu '1': specify which gpu to use
• --onegreen: if set the flag, sl_onegreen worker will start to train data crawled from game.onegreen.net
• --skip SKIP: if set this flag, games whoses index is less than SKIP would not be used to train (only valid when onegreen flag is set)