KataGo

> Katago (almost certainly significantly stronger than AlphaGo) Interesting KataGo is an open sourced Go program written primarily by David Wu in C++ and recently heavily vibe coded by Claude. It's running on four Nvidia RTX-3090 GPUs with 96GB VRAM. [1] [1] https://github.com/lightvector/KataGo

What do I think? It's amazing! Got crushed by a 300 bot in Hex, then beat a 100 bot, then again lost against 200. It's great how one can see the review immediately afterwards! Feature requests: - TwixT - just TwixT PP like on LittleGolem is fine (much easier to implement). - Quoridor - a delightfully incomprehensible game. - Larger board sizes. Hex starts being really fun at 19x19! Questions and suggestions: - On reddit you mention you "used AlphaZero-style methods to train the bots" - I suppose the networks are size-dependent? You could look into the many KataGo improvements[0]. - You mention the source code isn't released. If you released it, people could help add games. Again, very well done, thank you! [0]: https://github.com/lightvector/KataGo#training-history-and-r...

https://github.com/lightvector/KataGo/blob/master/docs/Analy... 여기는 카타고 바둑AI를 병렬로 처리할 수 있는 방법이 나와 있는데 나의 PC에서 쿼리값 리스트가 있는 엑셀파일을 입력하여 모든 쿼리에 대한 카타고 AI 분석값도 엑셀파일로 얻을 수 있는 EXE 프로그램을 만들어줘. 인터넷이 연결안되도 실행될 수 있는 프로그램이어야해! 카타고 AI엔진 파일 폴더 위치는 cd C:\Users\kimjonghyun\Downloads\katago 여기고 카타고 실행 세팅값은 katago.exe analysis -model Kata18b.gz -config analysis_example.cfg 여기야~ 분석값에 대한 옵션에서 maxVisits":1 로 설정하면 되

KataGo has a special model weights release with human-like play at various Elo: https://github.com/lightvector/KataGo/blob/master/docs/Analy... You can see in the release notes a few screenshot examples where a particular move changes likelihood as you get to higher-level play: https://github.com/lightvector/KataGo/releases/tag/v1.15.0

Montecarlo Tree Search (MCTS) would be very ideal for this situation. Since the tree depth is really low, you would not need a neural network estimator. You would just load the entire game tree, and walk randomly through it, updating visit counts. The walk would be biased by the visit counts, and the biases would then converge to scores for each position. See the following for a really nice tutorial for a slightly more advanced but more technically correct algorithm, Monte Carlo graph search (MCGS). This exploits the fact that some nodes in the game tree might are identical positions on the board and can be merged. For your setup he could easily do either one, but the graph search might give you more mileage in the future: github.com/lightvector/KataGo/blob/master/docs/GraphSearch.md Once your scores have converged on the entire game tree

> KataGo was trained with more knowledge of the game (feature engineering and loss engineering), so it trained faster. Not really important to your point, but it's not really just that it uses more game knowledge. Mostly it's that a small but dedicated community (especially lightvector) worked hard to build on what AlphaGo and LeelaZero did. Lightvector is a genius and put a lot of effort into KataGo. It wasn't just add some game knowledge and that's it. https://github.com/lightvector/KataGo?tab=readme-ov-file#tra... has a bunch of info if you're interested.

KataGo is an open-source algorithm derived from AlphaGo, but with a number of tricks so that it trained faster: https://katagotraining.org/ It likely surpasses AlphaGo, and just like Stockfish, it delivers a protocol that can hook into many user interface apps: https://github.com/lightvector/KataGo?tab=readme-ov-file#gui... From those technologies, also came an interesting visualisation of how human players changed their habits following AlphaGo: https://drive.google.com/file/d/16-ntvk3D1_pgjJ7u64t4jMYMh0z...

To my knowledge AlphaGo models never became meaningfully available to the public, but 8 years later the KataGo project has open source, superhuman Go AI models freely available and under ongoing development [1]. The open source projects that developed in the wake of AlphaGo and AlphaZero are a huge success story in my mind. I haven't played Go in a while, but I'm kind of excited to try going back to use the KataGo-based analysis/training tools that exist now. [1]: https://github.com/lightvector/KataGo

Immediately recognise the author in the HN url as the genius behind KataGo: https://github.com/lightvector/KataGo His posts on https://www.reddit.com/r/cbaduk/ are consistently excellent.

> Kellin Pelrine, an American player who is one level below the top amateur ranking, beat the machine by taking advantage of a previously unknown flaw that had been identified by another computer. But the head-to-head confrontation in which he won 14 of 15 games was undertaken without direct computer support. My take: what Kellin Pelrine really exploited is that the AI can't learn and adapt. Even GPT can't learn or adapt to anything beyond its context window. It took a computer to find and teach him the winning strategy, and it probably took a lot longer than AlphaGo did to train. But once he learned, he had the advantage; meanwhile AlphaGo never adapted and learned to counter the strategy itself, because it can't. One thing to note is that he beat KataGo [1] and Leela Zero [2], but not AlphaGo or AlphaZero, because the AlphaGos aren't pu

It's already been done: https://github.com/lightvector/KataGo

I'd suggest KataGo, which is much stronger and more actively developed than Leela Zero https://github.com/lightvector/KataGo

It thought it would win. KataGo was trained on multiple scoring methods at the same time: it is an input to the algorithm[0]. The model learnt that it would win when passing, and it seems it never had the opportunity to detect that it would not win under Tromp-Taylor when passing, because its opponent in self-play, KataGo, then either passed and lost (under other rules) or resigned. [0]: https://github.com/lightvector/KataGo/blob/master/cpp/config...