What is this?
This is a baduk (go) / Janggi engine running on the cloud. To be specific, this is a customized version of leela-zero (http://zero.sjeng.org/) running on a couple of consumer-grade GPUs, with a web frontend written from scratch.
To be more specific, this is a 192×15 net running at -v 800 to -v 2000 depending on server load, with some patches for performance optimization on this specific environment. This probably won’t be as good as people using their own GPUs, but it should be much faster than any consumer-grade CPU-only setups.
(Updated) : I developed a Janggi AI from scratch! This is heavily influenced from the leela-zero project using the same “AlphaZero” method – combination of reinforcement learning and monte-carlo tree search methods. This took 6 month of training from four consumer-grade GPUs, which is now playing at least at the level of very strong amateurs (and it is getting stronger and stronger)
What are the game settings here?
Baduk: Tromp-Taylor rules. 7.5 komi for even games. Handicap games run with 0.5 komi.
Janggi: Modern rule (“Kakao-Janggi” rule) – no draws, no repetitive moves, and loses once you go lower than 10 points.
What is stored on the server?
At this point, nothing related to the game itself is stored on the servers. Once you close your window, it’s all gone.
What’s your next plans?
- Some way to review the games you played – this is quite compute-intensive and I am thinking whether there is a good way to reuse your play record, but no good idea yet.