Page 4 of 8
Re: The new Tafl engine (AI)
Posted: Mon Apr 27, 2026 2:22 am
by sovaz1997
On 50K nodes (Nath vs. Bainn)
https://playtafl.org/?game=RwAzYWFhYWEz ... UWFQHgCm0B
It's equal to 10 minutes per move on the
https://playtafl.org/
Code: Select all
[1/28] black f2i2 winrate 41.8%
[2/28] white e5e2 winrate 39.3%
[3/28] black b6b3 winrate 43.3%
[4/28] white f5b5 winrate 37.1%
[5/28] black d1d2 winrate 48.5%
[6/28] white f4c4 winrate 47.0%
[7/28] black d2c2 winrate 51.5%
[8/28] white d6d1 winrate 47.8%
[9/28] black f10c10 winrate 67.2%
[10/28] white f6f3 winrate 68.1%
[11/28] black j6j3 winrate 74.4%
[12/28] white f3d3 winrate 73.6%
[13/28] black d11d4 winrate 88.9%
[14/28] white e6e4 winrate 90.1%
[15/28] black a6d6 winrate 91.1%
[16/28] white d1c1 winrate 91.3%
[17/28] black j3e3 winrate 92.6%
[18/28] white c4c3 winrate 92.8%
[19/28] black a4a2 winrate 93.9%
[20/28] white c3c6 winrate 94.1%
[21/28] black d6d5 winrate 95.6%
[22/28] white c6c5 winrate 94.0%
[23/28] black a7d7 winrate 96.2%
[24/28] white d3d1 winrate 94.9%
[25/28] black d5d2 winrate 99.5%
[26/28] white c1c4 winrate 99.5%
[27/28] black b3b1 winrate 99.6%
[28/28] white d1c1 winrate 99.8%
White accuracy: 98.7% | Black accuracy: 94.5%
@Draganov I gave some deep analysis examples.
Perhaps the engine analyzes well enough if given time.
I'm interested in your opinion on the proposed moves:
So, for engine, mistakes was in next moves:
5. f10c10 (suggested best: f1f3, c2c1, g1g3 later and even with small advantage for black)
6. j6j3 (suggested: h1h3). But position already not good for black
7. d11d4 (suggested: d11d6). After this "mistake" engine thinks that position probably lost for black with 10% chance for white estimated
Re: The new Tafl engine (AI)
Posted: Mon Apr 27, 2026 7:21 am
by Ytreza
Moving 2nd line attackers toward corners too early is usually a bad idea.
Playing against the engine I notice it loves playing on the 3rd row, even if it means letting defenders access corners.
I'm actually impressed how the engine playing white finds potential in the opening. It knows the right timing to get the king out, and when and how to invade a corner efficiently. In actual corner fights it might start doing weird things, like repetitions that don't make sense, or missing obvious tactics. And in endgame it's pretty bad at the moment.
To make it learn edge forts more efficiently, maybe you could start by training it on endgame problems instead of just letting it play against itself? Or do you think it would overfit?
Personnally I believe it is not strong enough yet to analyze top player games in a meaningful way. However I would very much like to see how it evaluates the game balance, even without edge forts. I'd suggest you take some models at various training stages and let them play eg 1000 games, then plot the balance as function of training stage. I can also try to do it but this week is busy
Re: The new Tafl engine (AI)
Posted: Mon Apr 27, 2026 9:20 am
by sovaz1997
Ytreza wrote: Mon Apr 27, 2026 7:21 am
Moving 2nd line attackers toward corners too early is usually a bad idea.
Playing against the engine I notice it loves playing on the 3rd row, even if it means letting defenders access corners.
I'm actually impressed how the engine playing white finds potential in the opening. It knows the right timing to get the king out, and when and how to invade a corner efficiently. In actual corner fights it might start doing weird things, like repetitions that don't make sense, or missing obvious tactics. And in endgame it's pretty bad at the moment.
To make it learn edge forts more efficiently, maybe you could start by training it on endgame problems instead of just letting it play against itself? Or do you think it would overfit?
Personnally I believe it is not strong enough yet to analyze top player games in a meaningful way. However I would very much like to see how it evaluates the game balance, even without edge forts. I'd suggest you take some models at various training stages and let them play eg 1000 games, then plot the balance as function of training stage. I can also try to do it but this week is busy
We already have the balance graphic, here for every 1 million positions in train process:
To make it learn edge forts more efficiently, maybe you could start by training it on endgame problems instead of just letting it play against itself? Or do you think it would overfit?
So far, ELO is growing thanks to basic training.
It's probably too early now, and manually feeding it data would be very difficult to provide enough. And yes, it can retrain itself from that.
Re: The new Tafl engine (AI)
Posted: Mon Apr 27, 2026 9:46 am
by Ytreza
We already have the balance graphic, here for every 1 million positions in train process:
Great! I assume the sharp transition is when it learned basic corner tactics?
If you need more ressources for training, I might have access to a supercomputing center.
Re: The new Tafl engine (AI)
Posted: Mon Apr 27, 2026 10:15 am
by sovaz1997
Definitely, it was corners
Re: The new Tafl engine (AI)
Posted: Mon Apr 27, 2026 10:26 am
by Ytreza
Actually white having 40% wins without edge forts is surprisingly good. I would have expected black to always be able to close corners.
Honestly when playing black against the engine I have a hard time defending corners and when I manage it's at the cost of a lot of pieces and leaving sides wide open. I get the impression the only reason I win is because the AI doesn't know the forts.
That said building edge forts is not an easy task so the game might actually end up being very balanced at the end of the day! Very exciting
Re: The new Tafl engine (AI)
Posted: Mon Apr 27, 2026 11:10 am
by sovaz1997
Yes, the result is good.
Maybe that's why the nn doesn't look at forts, since its existing knowledge is enough for it to create wins.
Re: The new Tafl engine (AI)
Posted: Mon Apr 27, 2026 11:11 am
by sovaz1997
Anyway I need ro check some things.
In some cases engine do 3-fold repetitions, but instead doing this it could continue playing differently.
Maybe this could improve training.
Re: The new Tafl engine (AI)
Posted: Mon Apr 27, 2026 11:48 am
by Ytreza
If white is about to repeat, could you prevent the move from being played at all?
You could also try to implement the "partial surround" rule: black wins if white cannot escape or build a fort anymore, even when given an arbitrary number of free moves. It's a bit annoying to check algorithmically but that would save computing resources by cutting lengthy endgame, and that could help white understanding edge forts faster.
In practice you can approximate this win check by doing flood-fill with white, then capture black pieces which become sandwiched, then flood-fill again, repeat until there are no capture possible, and finally remove all white pieces and check if the king can access a corner.
Re: The new Tafl engine (AI)
Posted: Mon Apr 27, 2026 11:49 am
by Ytreza
Another interesting stat is the number of pieces remaining (especially black pieces) at the end of the game. If white manages to capture more than 6 in corner fights, it becomes difficult for black to later defend against forts.