Tensorflow仿AlphaGo框架实现的AI围棋程序

Tensorflow仿AlphaGo框架实现的AI围棋程序

JoshieGo

A Go playing program implemented in Tensorflow roughly according to the architecture of AlphaGo. JoshieGo has strong policy/value network, and uses APV-MCTS to search during live play. On a single GTX1080 given 20 seconds per move, JoshieGo achieve strength around 3~4 amateur dan on OGS and Tencent's Fox Go Server.

What is special about JoshieGo?

Compared to many other AlphaGo-related but failed projects, JoshieGo does not aim to faithfully replicate AlphaGo's original paper, especially in the training of value network which is crucial for final performance. Rather, JoshieGo aims to build practical and strong Go engine with acceptable time and affordable hardware. The value network of JoshieGo is trained and fine-tuned purely with publicly available data using supervised learning without obvious overfitting. Policy network and MCTS are implemented as described in AlphaGo paper, with the exception that we do not use a fast roll-out policy.

How to play

Requiments

Tensorflow == 1.0.0

Numpy >= 1.11.1

OpenCV (optional, visualization use only)

Start a Game

Edit IP Address in play.py run python play.py --is_server=1 run python play.py

Games by JoshieGo

Games played by JoshieGo can be found by searching username "JoshieBot" on Tencent's Fox Go Server.

Futher work

Surprisingly, JoshieGo can not read ladder. Feature planes indicating successful ladder capture and ladder escape will be added.Wrap feature extraction code in C to boost performance.Implement distributed APV-MCTS.

版权声明：本文内容由网络用户投稿，版权归原作者所有，本站不拥有其著作权，亦不承担相应法律责任。如果您发现本站中有涉嫌抄袭或描述失实的内容，请联系我们jiasou666@gmail.com 处理，核实后本网站将在24小时内删除侵权内容。