ArnoHu wrote:

Kaufman Test with current GoK Classic:

Problem    Found    Nodecount (k)    Comment
 1:          Y              4,154
 2:          Y                155
 3:          Y                  7    Lost again at 101k
 4:          Y              1,127
 5:          Y                248
 6:          Y                  0    Lost again at 22k
 7:          Y                253
 8:          Y               1341
 9:          Y                397
10:          Y                 40
11:          Y                131
12:          
13:          Y                618
14:          Y              1,859
15:          Y                285
16:          Y                106
17:          Y                167
18:          
19:          Y                 59
20:          Y                565
21:          
22:          Y                  2
23:          Y                  1    Lost again at 647k
24:          Y                 52
25:          

waabooboo wrote:

ArnoHu wrote:

Kaufman Test with current GoK Classic:

Problem    Found    Nodecount (k)    Comment
 1:          Y              4,154
 2:          Y                155
 3:          Y                  7    Lost again at 101k
 4:          Y              1,127
 5:          Y                248
 6:          Y                  0    Lost again at 22k
 7:          Y                253
 8:          Y               1341
 9:          Y                397
10:          Y                 40
11:          Y                131
12:          
13:          Y                618
14:          Y              1,859
15:          Y                285
16:          Y                106
17:          Y                167
18:          
19:          Y                 59
20:          Y                565
21:          
22:          Y                  2
23:          Y                  1    Lost again at 647k
24:          Y                 52
25:          

Nice! 21-25 are really more about endgame knowledge than tactical strength. 25 seems particularly tough for HCE engines – the king and pawn ending is clearly winning, but the engines have to calculate very deeply before trading for the “weak” minor piece.

21 is a bit broken, looks like it's a tablebase win for black no matter what. But any sane human player takes the knight, because you have much better practical drawing chances with the bishop endgame.

New Wolverine version released today – I got LMP working and retuned the search parameters. I'll run it through the gauntlet, see if the 17/25 can be improved upon

waabooboo wrote:

ArnoHu wrote:

Kaufman Test with current GoK Classic:

Problem    Found    Nodecount (k)    Comment
 1:          Y              4,154
 2:          Y                155
 3:          Y                  7    Lost again at 101k
 4:          Y              1,127
 5:          Y                248
 6:          Y                  0    Lost again at 22k
 7:          Y                253
 8:          Y               1341
 9:          Y                397
10:          Y                 40
11:          Y                131
12:          
13:          Y                618
14:          Y              1,859
15:          Y                285
16:          Y                106
17:          Y                167
18:          
19:          Y                 59
20:          Y                565
21:          
22:          Y                  2
23:          Y                  1    Lost again at 647k
24:          Y                 52
25:          

Nice! 21-25 are really more about endgame knowledge than tactical strength. 25 seems particularly tough for HCE engines – the king and pawn ending is clearly winning, but the engines have to calculate very deeply before trading for the “weak” minor piece.

21 is a bit broken, looks like it's a tablebase win for black no matter what. But any sane human player takes the knight, because you have much better practical drawing chances with the bishop endgame.

New Wolverine version released today – I got LMP working and retuned the search parameters. I'll run it through the gauntlet, see if the 17/25 can be improved upon

ArnoHu wrote:

waabooboo wrote:

ArnoHu wrote:

Kaufman Test with current GoK Classic:

Problem    Found    Nodecount (k)    Comment
 1:          Y              4,154
 2:          Y                155
 3:          Y                  7    Lost again at 101k
 4:          Y              1,127
 5:          Y                248
 6:          Y                  0    Lost again at 22k
 7:          Y                253
 8:          Y               1341
 9:          Y                397
10:          Y                 40
11:          Y                131
12:          
13:          Y                618
14:          Y              1,859
15:          Y                285
16:          Y                106
17:          Y                167
18:          
19:          Y                 59
20:          Y                565
21:          
22:          Y                  2
23:          Y                  1    Lost again at 647k
24:          Y                 52
25:          

Nice! 21-25 are really more about endgame knowledge than tactical strength. 25 seems particularly tough for HCE engines – the king and pawn ending is clearly winning, but the engines have to calculate very deeply before trading for the “weak” minor piece.

21 is a bit broken, looks like it's a tablebase win for black no matter what. But any sane human player takes the knight, because you have much better practical drawing chances with the bishop endgame.

New Wolverine version released today – I got LMP working and retuned the search parameters. I'll run it through the gauntlet, see if the 17/25 can be improved upon

Congrats to the new Wolverine version, it looks very strong.

The previous GoK Classic was lucky to draw these two games:

https://lichess.org/study/SaWdnTo2/ye99xDJU
https://lichess.org/study/SaWdnTo2/P2WeK47m

I then improved and reactivated internal iterative deepening and internal iterative reductions, and Classic managed to win on the same opening, still quite a fight at 91% vs. 88%:

https://lichess.org/study/SaWdnTo2/o4Z1KWVx

And after two evaluation bugfixes, 95% vs. 91%:

https://lichess.org/study/SaWdnTo2/cY0AaaSI

waabooboo wrote:

ArnoHu wrote:

waabooboo wrote:

ArnoHu wrote:

Kaufman Test with current GoK Classic:

Problem    Found    Nodecount (k)    Comment
 1:          Y              4,154
 2:          Y                155
 3:          Y                  7    Lost again at 101k
 4:          Y              1,127
 5:          Y                248
 6:          Y                  0    Lost again at 22k
 7:          Y                253
 8:          Y               1341
 9:          Y                397
10:          Y                 40
11:          Y                131
12:          
13:          Y                618
14:          Y              1,859
15:          Y                285
16:          Y                106
17:          Y                167
18:          
19:          Y                 59
20:          Y                565
21:          
22:          Y                  2
23:          Y                  1    Lost again at 647k
24:          Y                 52
25:          

Nice! 21-25 are really more about endgame knowledge than tactical strength. 25 seems particularly tough for HCE engines – the king and pawn ending is clearly winning, but the engines have to calculate very deeply before trading for the “weak” minor piece.

21 is a bit broken, looks like it's a tablebase win for black no matter what. But any sane human player takes the knight, because you have much better practical drawing chances with the bishop endgame.

New Wolverine version released today – I got LMP working and retuned the search parameters. I'll run it through the gauntlet, see if the 17/25 can be improved upon

Congrats to the new Wolverine version, it looks very strong.

The previous GoK Classic was lucky to draw these two games:

https://lichess.org/study/SaWdnTo2/ye99xDJU
https://lichess.org/study/SaWdnTo2/P2WeK47m

I then improved and reactivated internal iterative deepening and internal iterative reductions, and Classic managed to win on the same opening, still quite a fight at 91% vs. 88%:

https://lichess.org/study/SaWdnTo2/o4Z1KWVx

And after two evaluation bugfixes, 95% vs. 91%:

https://lichess.org/study/SaWdnTo2/cY0AaaSI

Thanks! I'll work on search a bit more – Wolverine still can't solve test position #19, which GoK got quickly. Probably a futility pruning thing, I'll take a look… And there's still work to do on the evaluation

ArnoHu wrote:

Connect 4 Bot Game Series

I am conducting a Connect 4 Bot test game series played by Game of Connect (GoC).

The results will be updated in this posting over time. They will cover the Scratch runtime and TurboWarp (TW only if an engine takes advantage of the additional performance there), with both sides as starting players once (as far as the projects allow). When several playing levels exist, games will be played on highest level available for the runtime.

The Bots

• Connect 4 - “GoC”: https://scratch.mit.edu/projects/1114744709/ (by ArnoHu)
  
• Connect Four AI: https://scratch.mit.edu/projects/1109505895 (by BirdRacerThree)
  
• Connect 4 AI (GreenLocked): https://scratch.mit.edu/projects/42395968 (by PullJosh)
  
• Connect 4 AI: https://scratch.mit.edu/projects/802293688 (by Spentine)
  
• Online Connect 4: https://scratch.mit.edu/projects/685271634 (by AnotherKidOnScratch)
  
• Connect 4 vs. Computer: https://scratch.mit.edu/projects/16243147 (by A_big_cow)
  

Scratch 3

Starting Side       Responding Side

GoC              -  CFAI (BirdRacer)     1  -  0    (win at ply 41)
GoC              -  C4AI (Spentine)      1  -  0    (win at ply 29)
GoC              -  C4AI-GL (PullJosh)   1  -  0    (win at ply 25)
GoC              -  OC4 (AKOS)           1  -  0    (win at ply 33)
GoC              -  OC4VC (ABC)          1  -  0    (win at ply 31)

CFAI (BirdRacer) -  GoC                0.5  -  0.5
C4AI (Spentine)  -  GoC                  0  -  1    (loss at ply 34)

TurboWarp

Starting Side       Responding Side

GoC              -  CFAI (BirdRacer)     1  -  0    (win at ply 31)
GoC              -  C4AI (Spentine)      1  -  0    (win at ply 39)

CFAI (BirdRacer) -  GoC                0.5  -  0.5
C4AI (Spentine)  -  GoC                  0  -  1    (loss at ply 42)

Note: GoC and BirdRacer's CFAI are think-time bound, hence not deterministic. They will play according to system speed (means, better on faster systems).

If you reply to this post, I suggest to link to it and remove the copied content, otherwise we will have outdated results copied over the forum. Here is the link to the original: https://scratch.mit.edu/discuss/post/8319884/

ArnoHu wrote:

GoK Classic has just been updated.

GoK Classic vs. White Dove, 93% vs. 90%:
https://lichess.org/study/SaWdnTo2/VIKihROw

GoK Classic vs. Wolverine 2, 96% vs. 91%:
https://lichess.org/study/SaWdnTo2/T1PXShRm

GoK Classic vs. Black Crow, 90% vs. 91%:
https://lichess.org/study/SaWdnTo2/kzkdF7Tx

GoK Classic vs. TurboKnight 5, 91% vs. 89%:
https://lichess.org/study/SaWdnTo2/hsiOa53c

GoK Classic vs. Wolverine 2, 93% vs. 87%:
https://lichess.org/study/SaWdnTo2/O9jpMC3t

GoK Classic vs. Delta, 97% vs. 89%:
https://lichess.org/study/SaWdnTo2/XQ2dgj4U

GoK Classic vs. Shallow Blue 3 NNUE, 93% vs. 90%:
https://lichess.org/study/SaWdnTo2/U1VwFQpS

GoK Classic vs. White Dove, 98% vs. 92%:
https://lichess.org/study/SaWdnTo2/N0Im9Cjc

Two improvements merged to GoK NNUE:

GoK NNUE vs. Black Crow, 98% vs. 94%:
https://lichess.org/study/SaWdnTo2/JUhj2fEe

coooolboy55 wrote:

ArnoHu wrote:

Connect 4 Bot Game Series

I am conducting a Connect 4 Bot test game series played by Game of Connect (GoC).

The results will be updated in this posting over time. They will cover the Scratch runtime and TurboWarp (TW only if an engine takes advantage of the additional performance there), with both sides as starting players once (as far as the projects allow). When several playing levels exist, games will be played on highest level available for the runtime.

The Bots

• Connect 4 - “GoC”: https://scratch.mit.edu/projects/1114744709/ (by ArnoHu)
  
• Connect Four AI: https://scratch.mit.edu/projects/1109505895 (by BirdRacerThree)
  
• Connect 4 AI (GreenLocked): https://scratch.mit.edu/projects/42395968 (by PullJosh)
  
• Connect 4 AI: https://scratch.mit.edu/projects/802293688 (by Spentine)
  
• Online Connect 4: https://scratch.mit.edu/projects/685271634 (by AnotherKidOnScratch)
  
• Connect 4 vs. Computer: https://scratch.mit.edu/projects/16243147 (by A_big_cow)
  

Scratch 3

Starting Side       Responding Side

GoC              -  CFAI (BirdRacer)     1  -  0    (win at ply 41)
GoC              -  C4AI (Spentine)      1  -  0    (win at ply 29)
GoC              -  C4AI-GL (PullJosh)   1  -  0    (win at ply 25)
GoC              -  OC4 (AKOS)           1  -  0    (win at ply 33)
GoC              -  OC4VC (ABC)          1  -  0    (win at ply 31)

CFAI (BirdRacer) -  GoC                0.5  -  0.5
C4AI (Spentine)  -  GoC                  0  -  1    (loss at ply 34)

TurboWarp

Starting Side       Responding Side

GoC              -  CFAI (BirdRacer)     1  -  0    (win at ply 31)
GoC              -  C4AI (Spentine)      1  -  0    (win at ply 39)

CFAI (BirdRacer) -  GoC                0.5  -  0.5
C4AI (Spentine)  -  GoC                  0  -  1    (loss at ply 42)

Note: GoC and BirdRacer's CFAI are think-time bound, hence not deterministic. They will play according to system speed (means, better on faster systems).

If you reply to this post, I suggest to link to it and remove the copied content, otherwise we will have outdated results copied over the forum. Here is the link to the original: https://scratch.mit.edu/discuss/post/8319884/

Will you be running another Connect 4 tournament?

ArnoHu wrote:

GoK Classic - Shallow Blue 3 NNUE, 89% vs. 86%:
https://lichess.org/study/PJTsaIt9/ZyFQaV9d

GoK Classic - Wolverine 2 NNUE, 86% vs. 83%:
https://lichess.org/study/PJTsaIt9/vVV4YDrr

coooolboy55 wrote:

I have stopped working on my Chess AI. I really don't like creating it.

coooolboy55 wrote:

I have stopped working on my Chess AI. I really don't like creating it.

when green flag clicked
wait (3) secs
wait (3) secs

when green flag clicked
change [ v] by (Happy)
[/s[scratchblocks]
cratchblocks]
change [ v] by (sad)
[/scratchblocks]

iceysnowman wrote:

coooolboy55 wrote:

I have stopped working on my Chess AI. I really don't like creating it.

Noooooo! Too bad, it could have ended up really well. I at least hope you continue to work on your Connect 4 bot, it's a really nice project and it's pretty decent.