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.
I love TwixT, I discovered the Bookshelf Games version of this recently (a friend of mine's parents had an old copy at their beach house). Also a big fan of Alex Randolph's Ricochet Robots.
Dieter Stein's games are also supposed to be wonderful abstracts (Urbino, Fendo, Tintas) though I haven't had a chance to play those yet.
Wow, beating a 100 bot is still impressive in my book.
Thanks for the feedback and suggestions!
Yes, the networks are size dependent right now. It's a great idea to copy-paste and then adapt the KataGo network architecture since it isn't size dependent and has been proven to reach superhuman strengths.
I'm about 1950 on littlegolem, but trying again, I just lost a couple of games in a row against the 100 bot.
A couple more questions/remarks:
- It seems everything is happening on the client? My CPU (I'm on a laptop without a real GPU) goes wild during analysis. But also I don't notice any big bag of neural-net-weights being downloaded. Mind sharing how it works?
- Care to share more about the networks? How long did you train the networks for and on what GPU? Circa how many params? Any and all details you'd like to share :)
- This Tumbleweed game is fun! This game seems somewhat inspired by both Hex and Go, and the author lives in Warsaw. Interestingly, I lived in Warsaw in 2020, am very active in the Go scene (4 dan), and at least know the names of the people who play Hex, and yet somehow never heard about Michał or Tumbleweed before...
Sans proper profiling, I would guess that the CPU going wild during analysis is due to a combination of 1. analysis is streamed live to the client in 20 simulation intervals 2. some post-processing on the client side 3. the fact that I am using a global context and reducer in React which causes the entire page to re-render each time an update happens.
The networks are simple Resnets with a value and policy head. It's 20 layers with 128 channels per layer. I trained for several days on 2x 4090s. However, recently I trained a few networks (Hex 14x14, Amazons 10x10, Breakthrough 8x8) on a GH200 and it was 2x faster, roughly 100 ckpts per 24 hours for Hex 14x14. I'm not sure about the number of parameters but the .pt and .ts files are on the order of 30-90 MB. There's definitely room for improvement using tricks like quantization during selfplay inference.
I'm very happy you like Tumbleweed! If you're curious there's a Tumbleweed community run by Michał (the creator) https://discord.com/invite/wu6Xdtt497 They are currently playing through their 2025 World Championship.
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...