Scratch Chess Engine - Game of Kings

ArnoHu

GoK NNUE vs. Black Crow, 98% vs. 93%:
https://lichess.org/study/nHWsefvf/6l8PIRIv

ArnoHu

New GoK Classic with improved pawn positional evaluation:

GoK Classic vs. Shallow Blue 3 Dev, 92% vs. 87%:
https://lichess.org/study/nHWsefvf/HpSVq06C

GoK Classic vs. Delta, 97% vs. 97%:
https://lichess.org/study/nHWsefvf/FfEGCaEb

GoK Classic vs. White Dove, 87% vs. 87%:
https://lichess.org/study/nHWsefvf/3O5bNGdi

GoK Classic vs. Wolverine 2, 97% vs. 91%:
https://lichess.org/study/nHWsefvf/vHc0zVZC

With an other minor update, those are the results:

GoK Classic vs. Shallow Blue 3 Dev, 93% vs. 90%:
https://lichess.org/study/nHWsefvf/qe3yMidA

GoK Classic vs. Delta, 94% vs. 90%:
https://lichess.org/study/nHWsefvf/qFoOsWGG

GoK Classic vs. White Dove, 96% vs. 92%:
https://lichess.org/study/nHWsefvf/10hI9fal

GoK Classic vs. Wolverine 2, 87% vs. 83%:
https://lichess.org/study/nHWsefvf/f5UCnF6y

Last edited by ArnoHu (Oct. 11, 2025 12:52:56)

ArnoHu

GoK Classic vs. White Dove, 86% vs. 83%:
https://lichess.org/study/nHWsefvf/yyOQiwMD

GoK Classic vs. Wolverine 2, 87% vs. 78%:
https://lichess.org/study/nHWsefvf/js4wVmzb

waabooboo

ArnoHu wrote:
GoK Classic vs. White Dove, 86% vs. 83%:
https://lichess.org/study/nHWsefvf/yyOQiwMD

GoK Classic vs. Wolverine 2, 87% vs. 78%:
https://lichess.org/study/nHWsefvf/js4wVmzb

Looks like futility pruning cost Wolverine this game… I currently allow it up to 4 plies from the horizon, so there's a bunch of lines where Wolverine prunes forward after rook takes queen, overlooking Rg8# on the next turn.

With futility pruning disabled, Wolverine senses the danger within the 5 second time limit and avoids the mate. But I think disabling this harms his overall quality of play. I'm curious how GoK handles futility pruning, if there's a better way to do it.

Edit: I guess what I have implemented is reverse futility pruning, not standard futility pruning. I check if static eval > beta + margin before looping through the moves, and if the condition is true then I fail high immediately.

Last edited by waabooboo (Oct. 13, 2025 11:47:36)

ArnoHu

waabooboo wrote:
ArnoHu wrote:
GoK Classic vs. White Dove, 86% vs. 83%:
https://lichess.org/study/nHWsefvf/yyOQiwMD

GoK Classic vs. Wolverine 2, 87% vs. 78%:
https://lichess.org/study/nHWsefvf/js4wVmzb

Looks like futility pruning cost Wolverine this game… I currently allow it up to 4 plies from the horizon, so there's a bunch of lines where Wolverine prunes forward after rook takes queen, overlooking Rg8# on the next turn.

With futility pruning disabled, Wolverine senses the danger within the 5 second time limit and avoids the mate. But I think disabling this harms his overall quality of play. I'm curious how GoK handles futility pruning, if there's a better way to do it.

Edit: I guess what I have implemented is reverse futility pruning, not standard futility pruning. I check if static eval > beta + margin before looping through the moves, and if the condition is true then I fail high immediately.

GoK Classic vs. Wolverine 2, 91% vs. 86%:
https://lichess.org/study/nHWsefvf/bbFXfCxI

GoK's game was better than it looks like, blunder 1 prevented in 9.8sec, blunders 2+3 were m7's it didnt see

Last edited by ArnoHu (Oct. 14, 2025 06:13:38)

ArnoHu

GoK Classic vs. White Dove, 97% vs. 92%:
https://lichess.org/study/nHWsefvf/yJZVK22z

GoK Classic vs. Delta, 89% vs. 89%:
https://lichess.org/study/nHWsefvf/yCpon3ai

GoK Classic vs. Shallow Blue 3 Dev, 92% vs. 85%:
https://lichess.org/study/nHWsefvf/37nDopPX

GoK Classic vs. Wolverine 2, 97% vs. 92%:
https://lichess.org/study/nHWsefvf/S58aEapL

Last edited by ArnoHu (Oct. 14, 2025 17:15:48)

ArnoHu

ArnoHu wrote:
GoK Classic vs. White Dove, 97% vs. 92%:
https://lichess.org/study/nHWsefvf/yJZVK22z

GoK Classic vs. Delta, 89% vs. 89%:
https://lichess.org/study/nHWsefvf/yCpon3ai

GoK Classic vs. Shallow Blue 3 Dev, 92% vs. 85%:
https://lichess.org/study/nHWsefvf/37nDopPX

GoK Classic vs. Wolverine 2, 97% vs. 92%:
https://lichess.org/study/nHWsefvf/S58aEapL

Just one more improvement on king ring attacker weights, and GoK Classic plays a perfect game against White Dove, 98% vs. 91%:
https://lichess.org/study/nHWsefvf/ZoDcrGj3

GoK Classic vs. Wolverine 2, 94% vs. 92%:
https://lichess.org/study/nHWsefvf/vgPLbV65

GoK Classic vs. Delta, 95% vs. 90%:
https://lichess.org/study/nHWsefvf/h39llBU8

Last edited by ArnoHu (Oct. 15, 2025 02:01:23)

ArnoHu

ArnoHu wrote:
ArnoHu wrote:
GoK Classic vs. White Dove, 97% vs. 92%:
https://lichess.org/study/nHWsefvf/yJZVK22z

GoK Classic vs. Delta, 89% vs. 89%:
https://lichess.org/study/nHWsefvf/yCpon3ai

GoK Classic vs. Shallow Blue 3 Dev, 92% vs. 85%:
https://lichess.org/study/nHWsefvf/37nDopPX

GoK Classic vs. Wolverine 2, 97% vs. 92%:
https://lichess.org/study/nHWsefvf/S58aEapL

Just one more improvement on king ring attacker weights, and GoK Classic plays a perfect game against White Dove, 98% vs. 91%:
https://lichess.org/study/nHWsefvf/ZoDcrGj3

GoK Classic vs. Wolverine 2, 94% vs. 92%:
https://lichess.org/study/nHWsefvf/vgPLbV65

GoK Classic vs. Delta, 95% vs. 90%:
https://lichess.org/study/nHWsefvf/h39llBU8

GoK Classic vs. Shallow Blue 3 Dev, 92% vs. 86%:
https://lichess.org/study/nHWsefvf/8iaXTSk6

GoK Classic vs. White Dove, 92% vs. 87%:
https://lichess.org/study/nHWsefvf/cJjjLc5d

GoK Classic vs. White Dove, 94% vs. 85%:
https://lichess.org/study/nHWsefvf/IpIRYoWO

GoK Classic vs. Delta, 93% vs. 88%:
https://lichess.org/study/nHWsefvf/3L33Lee0

coooolboy55

What activations do you use for your NNUEs?

Destructor_chess

coooolboy55 wrote:
What activations do you use for your NNUEs?

reLu I think

ArnoHu

Destructor_chess wrote:
coooolboy55 wrote:
What activations do you use for your NNUEs?
reLu I think

Correct!

ArnoHu

GoK Classic vs. White Dove, 96% vs. 90%:
https://lichess.org/study/7cqWfi6q/rhnmquUT

GoK Classic vs. Delta, 85% vs. 80%:
https://lichess.org/study/7cqWfi6q/Tf1LHjLF

GoK Classic vs. Shallow Blue 3 Dev, 94% vs. 93%:
https://lichess.org/study/7cqWfi6q/7rLDSYiv

GoK Classic vs. Wolverine 2, 98% vs. 97%:
https://lichess.org/study/7cqWfi6q/WpKjyJ9D

GoK Classic vs. Wolverine 2, 87% vs. 83%:
https://lichess.org/study/7cqWfi6q/tb85ykvi

GoK Classic vs. Shallow Blue 3 Dev, 92% vs. 88%:
https://lichess.org/study/7cqWfi6q/vzySbbzJ

Last edited by ArnoHu (Oct. 16, 2025 10:01:58)

WhiteDoveOfficial

coooolboy55 wrote:
What activations do you use for your NNUEs?

BC uses ScReLU as generally speaking it gives slightly better results compared to standard ReLU.

Destructor_chess

WhiteDoveOfficial wrote:
coooolboy55 wrote:
What activations do you use for your NNUEs?

BC uses ScReLU as generally speaking it gives slightly better results compared to standard ReLU.

But the training is more noisy I think, and it takes more ressources.

waabooboo

New Wolverine version out, with some more improvements to the transposition table. This helps endgame search depth quite a bit.

https://imgur.com/a/fZhYKqb – I think Wolverine will not lose endgames like this any more

ArnoHu

waabooboo wrote:
New Wolverine version out, with some more improvements to the transposition table. This helps endgame search depth quite a bit.

https://imgur.com/a/fZhYKqb – I think Wolverine will not lose endgames like this any more

Plays pretty well, esp. midgame:
https://lichess.org/study/7cqWfi6q/NjTsbPil

Destructor_chess

Do you store checkmates in TT wabooboo?

Destructor_chess

Wich think time for this depth? 5 sec?

waabooboo wrote:
New Wolverine version out, with some more improvements to the transposition table. This helps endgame search depth quite a bit.

https://imgur.com/a/fZhYKqb – I think Wolverine will not lose endgames like this any more

10?

waabooboo

Destructor_chess wrote:
Wich think time for this depth? 5 sec?
waabooboo wrote:
New Wolverine version out, with some more improvements to the transposition table. This helps endgame search depth quite a bit.

https://imgur.com/a/fZhYKqb – I think Wolverine will not lose endgames like this any more
10?

5 seconds, yes. Still around 200k nps, but it's a little slower than before because I generate moves before applying TT evals (just to ensure legality). For now, known wins from TT are +15000 and actual mates are 100000-ply.

waabooboo

ArnoHu wrote:
waabooboo wrote:
New Wolverine version out, with some more improvements to the transposition table. This helps endgame search depth quite a bit.

https://imgur.com/a/fZhYKqb – I think Wolverine will not lose endgames like this any more

Plays pretty well, esp. midgame:
https://lichess.org/study/7cqWfi6q/NjTsbPil

Nice, good to see! Midgame was back and forth there, looks like Wolverine had several chances to hold. But I can't blame him, that position is really complicated.

Discuss Scratch