What is IaGo?

IaGo is an Othello AI using SL(supervised learning) policy network, value network, and MCTS(Monte Carlo tree search) inspired by AlphaGo.
Short description in English:
IaGo: an Othello AI inspired by AlphaGo
Description in Japanese:
AlphaGoを模したオセロAIを作る(1): SLポリシーネットワーク - Qiita
AlphaGoを模したオセロAIを作る(2): RLポリシーネットワーク - Qiita
AlphaGoを模したオセロAIを作る(3): バリューネットワーク - Qiita
AlphaGoを模したオセロAIを作る(4): モンテカルロ木探索 - Qiita

How to play?

1. Install chainer
   $ pip install chainer

2. Download this repository
   $ git clone [email protected]:shionhonda/IaGo.git

3. Move to IaGo directory and execute game.py
   $ python game.py
   You can set following options:
   --auto=False or --a=False
   If this is set True, autoplay begins between SLPolicy and PV-MCTS, and if False (default), the game is played by you and PV-MCTS.
   The thinking time is 10 seconds.

4. When placing a stone, input two numbers separated by comma. For example:
   4,3
   The first number corresponds to the vertical position and the second to the horizontal (one origin).

How to train?

1. Download data from http://meipuru-344.hatenablog.com/entry/2017/11/27/205448
2. Save it as "IaGo/data/data.txt"
3. Augment data
   $ python load.py
   You need at least 32MB RAM to complete this step.
4. Execute train_policy.py to train SL policy network.
   $ python train_policy.py --policy=sl --epoch=10 --gpu=0
   You need GPUs to complete this step. It will take about 12 hours.
5. Execute train_policy.py to train rollout policy.
   $ python train_policy.py --policy=rollout --epoch=1 --gpu=0
   This is fast.
6. Execute train_rl.py to reinforce SL policy network with REINFORCE (a kind of policy gradients).
   $ python train_rl.py --set=10000
7. Execute train_value.py to train value network.
   $ python train_value.py --epoch=20 --gpu=0
8. Training done!

Acknowledgements

Special thanks to:
@Rochestar-NRT for replication of AlphaGo (especially MCTS).
@lazmond3 for giving lots of feedbacks!