Scratch Chess Engine - Game of Kings

internet44

found out that the pawn structure code in shallow blue didn't work properly, I had left some custom block operators blank by accident. with that fixed, here is an updated 5s game against Wolverine 2, if you'll remember this was a draw previously: https://lichess.org/cIkURCzA

Destructor_chess

sb broke tk v2 lol https://lichess.org/aGIehRGR#0

ArnoHu

ArnoHu wrote:
WhiteDoveOfficial wrote:
ArnoHu wrote:
The next GoK NNUE version (currently on dev stage at https://turbowarp.org/1033379869) looks pretty strong, 97% vs. 88% against Black Crow:
https://lichess.org/study/4JLzgBmK/JujDSpV7

GoK's only inaccuracy was not detected by BC.

Looks like the super aggressive net is both a blessing and a curse lol. I might have to add a slight HCE term for S3, to try to add a penalty for being down material, as I cannot trust BC on S3 to actually successfully complete king attacks, where's it's basically just half-blind

GoK NNUE vs. Black Crow, 98% vs. 91%:
https://lichess.org/study/4JLzgBmK/bqo1tE7G

GoK Classic vs. Shallow Blue 3, 90% vs. 86%:
https://lichess.org/study/4JLzgBmK/YFKxvXqJ

waabooboo

Are pawn storm/shield middlegame bonuses applied at the start of the game, or only after castling?

The implementations I looked at seem to apply these values from move 1. But I tried this in Wolverine, and it harmed his quality of play. With the pawn shield bonuses, Wolverine plays on the flanks, not castling or moving the central pawns. Not sure if my implementation is correct

coooolboy55

I made my check detection faster for my Connect 4 AI. Currently, it doesn't help its play much (except on scratch), but it will help me add more to it (like bias for the random games), to make it better. The other times I added bias to the random move generator, it made the AI worse for some reason. I suspect it was because it slowed the AI down, resulting in less accurate results.

grkw2020

coooolboy55 wrote:
I made my check detection faster for my Connect 4 AI. Currently, it doesn't help its play much (except on scratch), but it will help me add more to it (like bias for the random games), to make it better. The other times I added bias to the random move generator, it made the AI worse for some reason. I suspect it was because it slowed the AI down, resulting in less accurate results.

Is the updated one in the project? Because that win detection can still be sped up a LOT.

coooolboy55

grkw2020 wrote:
coooolboy55 wrote:
I made my check detection faster for my Connect 4 AI. Currently, it doesn't help its play much (except on scratch), but it will help me add more to it (like bias for the random games), to make it better. The other times I added bias to the random move generator, it made the AI worse for some reason. I suspect it was because it slowed the AI down, resulting in less accurate results.
Is the updated one in the project? Because that win detection can still be sped up a LOT.

Yes.

grkw2020

coooolboy55 wrote:
grkw2020 wrote:
coooolboy55 wrote:
I made my check detection faster for my Connect 4 AI. Currently, it doesn't help its play much (except on scratch), but it will help me add more to it (like bias for the random games), to make it better. The other times I added bias to the random move generator, it made the AI worse for some reason. I suspect it was because it slowed the AI down, resulting in less accurate results.
Is the updated one in the project? Because that win detection can still be sped up a LOT.
Yes.

If you want to make it even faster, start from where the last piece was inserted. All possible rows of four will contain that piece.

coooolboy55

grkw2020 wrote:
coooolboy55 wrote:
grkw2020 wrote:
coooolboy55 wrote:
I made my check detection faster for my Connect 4 AI. Currently, it doesn't help its play much (except on scratch), but it will help me add more to it (like bias for the random games), to make it better. The other times I added bias to the random move generator, it made the AI worse for some reason. I suspect it was because it slowed the AI down, resulting in less accurate results.
Is the updated one in the project? Because that win detection can still be sped up a LOT.
Yes.
If you want to make it even faster, start from where the last piece was inserted. All possible rows of four will contain that piece.

That is almost exactly what I am doing.

grkw2020

coooolboy55 wrote:
grkw2020 wrote:
coooolboy55 wrote:
grkw2020 wrote:
coooolboy55 wrote:
I made my check detection faster for my Connect 4 AI. Currently, it doesn't help its play much (except on scratch), but it will help me add more to it (like bias for the random games), to make it better. The other times I added bias to the random move generator, it made the AI worse for some reason. I suspect it was because it slowed the AI down, resulting in less accurate results.
Is the updated one in the project? Because that win detection can still be sped up a LOT.
Yes.
If you want to make it even faster, start from where the last piece was inserted. All possible rows of four will contain that piece.
That is almost exactly what I am doing.

Really? Huh, it looks quite different to my implementation.

Destructor_chess

coooolboy55 wrote:
grkw2020 wrote:
coooolboy55 wrote:
grkw2020 wrote:
coooolboy55 wrote:
I made my check detection faster for my Connect 4 AI. Currently, it doesn't help its play much (except on scratch), but it will help me add more to it (like bias for the random games), to make it better. The other times I added bias to the random move generator, it made the AI worse for some reason. I suspect it was because it slowed the AI down, resulting in less accurate results.
Is the updated one in the project? Because that win detection can still be sped up a LOT.
Yes.
If you want to make it even faster, start from where the last piece was inserted. All possible rows of four will contain that piece.
That is almost exactly what I am doing.

You can also search in 2 directions at same time, or with look-up tables…

grkw2020

Destructor_chess wrote:
coooolboy55 wrote:
grkw2020 wrote:
coooolboy55 wrote:
grkw2020 wrote:
coooolboy55 wrote:
I made my check detection faster for my Connect 4 AI. Currently, it doesn't help its play much (except on scratch), but it will help me add more to it (like bias for the random games), to make it better. The other times I added bias to the random move generator, it made the AI worse for some reason. I suspect it was because it slowed the AI down, resulting in less accurate results.
Is the updated one in the project? Because that win detection can still be sped up a LOT.
Yes.
If you want to make it even faster, start from where the last piece was inserted. All possible rows of four will contain that piece.
That is almost exactly what I am doing.
You can also search in 2 directions at same time, or with look-up tables…

Wait, you're still here? Also what do you mean by looking in two directions at the same time. And wouldn't look up tables not really help without bitwise operators?

Last edited by grkw2020 (Sept. 16, 2025 08:04:33)

coooolboy55

I gave my bot the capability of playing Connect 4 at high speeds. I switched from the scratch timer to the “days since 2000” block (which is more accurate). It's quite funny that my bot can beat/tie some Connect 4 bots at the speed of 0.1 seconds.

grkw2020

coooolboy55 wrote:
I gave my bot the capability of playing Connect 4 at high speeds. I switched from the scratch timer to the “days since 2000” block (which is more accurate). It's quite funny that my bot can beat/tie some Connect 4 bots at the speed of 0.1 seconds.

Impressive!

Destructor_chess

coooolboy55 wrote:
I gave my bot the capability of playing Connect 4 at high speeds. I switched from the scratch timer to the “days since 2000” block (which is more accurate). It's quite funny that my bot can beat/tie some Connect 4 bots at the speed of 0.1 seconds.

How much knps?

WhiteDoveOfficial

coooolboy55 wrote:
I gave my bot the capability of playing Connect 4 at high speeds. I switched from the scratch timer to the “days since 2000” block (which is more accurate). It's quite funny that my bot can beat/tie some Connect 4 bots at the speed of 0.1 seconds.

Nice!

(BTW, have you tried a few games vs GoC or EosZero?)

coooolboy55

WhiteDoveOfficial wrote:
coooolboy55 wrote:
I gave my bot the capability of playing Connect 4 at high speeds. I switched from the scratch timer to the “days since 2000” block (which is more accurate). It's quite funny that my bot can beat/tie some Connect 4 bots at the speed of 0.1 seconds.

Nice!

(BTW, have you tried a few games vs GoC or EosZero?)

My Connect 4 AI, at normal computing time, cannot beat the best Connect 4 engines, also at normal computing time. I tried to put GoC on 0.1 computing time (by changing the code), but for some reason, the computing time wasn't very accurate. I was able to find out how to put computing time on EosZero though. Here is a game of EosZero vs Connect4AI (mine) at 0.1 seconds per move, with mine starting: 434444374333377711111177666666222222555555

Suprisingly it drew against my AI.

EosZero vs Connect4AI at 0.1 seconds per move, with yours starting: 455255544442222777711237745111133

A win against my AI.

S_P_A_R_T_Test

coooolboy55 wrote:
WhiteDoveOfficial wrote:
coooolboy55 wrote:
I gave my bot the capability of playing Connect 4 at high speeds. I switched from the scratch timer to the “days since 2000” block (which is more accurate). It's quite funny that my bot can beat/tie some Connect 4 bots at the speed of 0.1 seconds.

Nice!

(BTW, have you tried a few games vs GoC or EosZero?)
My Connect 4 AI, at normal computing time, cannot beat the best Connect 4 engines, also at normal computing time. I tried to put GoC on 0.1 computing time (by changing the code), but for some reason, the computing time wasn't very accurate. I was able to find out how to put computing time on EosZero though. Here is a game of EosZero vs Connect4AI (mine) at 0.1 seconds per move, with mine starting: 434444374333377711111177666666222222555555

Suprisingly it drew against my AI.

EosZero vs Connect4AI at 0.1 seconds per move, with yours starting: 455255544442222777711237745111133

A win against my AI.

Thanks for the games!

Technically, Connect 4 is a theoretical win for starting player, so a draw is still quite a good result for going second

coooolboy55

I updated my Connect 4 AI with move bias. It actually stands a chance against GoC.

coooolboy55

coooolboy55 wrote:
I updated my Connect 4 AI with move bias. It actually stands a chance against GoC.

Now compute time actually matters since the move bias, while making it better, also affects the amount of games played to accurately determine the best move.

Discuss Scratch