search

jhonnold / berserk

UCI Chess Engine written in C
GNU General Public License v3.0
200 stars 30 forks source link

Berserk 11 Critical Blunder #488

Closed ginkgo20 closed 1 year ago

ginkgo20 commented 1 year ago

Screenshot 2023-06-07 at 14-45-53 Analysis Board - 365Chess

8/4b2k/2b3pp/1p2p3/p3N1PP/P1N1P1K1/1P6/8 w - - 0 56

tissatussa commented 1 year ago

Can you explain ? Berserk was playing White or Black ? And what was the blunder ?

ginkgo20 commented 1 year ago

Berserk was playing White and h4h5 is blunder

tissatussa commented 1 year ago

h4h5 is blunder

Why do you state that ? What is your reference ? Mine is the newest ShashChess GZ NNUE, which tells me Black is better in this position (about -1) and White best moves are g5 and h5 !?

ginkgo20 commented 1 year ago

My reference is Stockfish 15 and Power Fritz 18. See this :

Untitled

tissatussa commented 1 year ago

minus 3.12 !? I just tested SF 15 NNUE and Komodo 14.1 and also SF CorChess v3 NN .. they all give g5 and h5 are best, although their eval differs (some engine show a 'conservative' eval).

is your screenshot from SF or Fritz ? Which depth ? Btw. i'm on Linux and i cannot use Fritz ..

tissatussa commented 1 year ago

i know what go depth 60 does : the engine will calculate all evals on all depths upto depth 60 and then give a best move .. when i let SF 15 run for 10 minutes it reaches depth 45 on my machine, and indeed g5 is a bit better than h5 .. not to say it's a "Critial Blunder".

maybe depth 60 will show -3.12 at last, but this will take a long time .. how did you ever come up with the idea to let your engine run upto depth 60 in this position ? Do it overnight ? When this position would ever occur in a real game, the engine will not have time to calculate upto depth 60, does it ?

berserk-critical-blunder-why

ginkgo20 commented 1 year ago

I did not go depth 60 , Also I let SF 15 run for 10 minutes same you but with this 2 comment :

position fen 8/4b2k/2b3pp/1p2p3/p3N1PP/P1N1P1K1/1P6/8 w - - 0 56 moves g4g5

position fen 8/4b2k/2b3pp/1p2p3/p3N1PP/P1N1P1K1/1P6/8 w - - 0 56 moves h4h5

If you want, I can create a challenge with you in Lichess via h4h5 in rapid game , I'm sure h4h5 is blunder ....

Edit : I think maybe there is a bug in "NN for the 2 Knights Position vs 2 Bishops Position" ?!

tissatussa commented 1 year ago

now tell me why it's a critical blunder ? I think with good play Black will win this position, because eval is about -1.6+ , but this will happen with both g5 and h5.

ginkgo20 commented 1 year ago

The position is equal. g5 is the only good move and h5 is blunder.

jhonnold commented 1 year ago

Why would you open an issue for this? Berserk is not perfect. No engine is. Making a mistake is simply going to happen.

ginkgo20 commented 1 year ago

@jhonnold Thanks for reply. I think there is a bug in NN training for "2 Knights Position vs 2 Bishops Position". Also similar to this issue found by @tissatussa for "4 Knights Position".

Screenshot 2023-06-07 at 14-45-53 Analysis Board - 365Chess

tissatussa commented 1 year ago

sorry, i see no relation with my "4 Knights Position".

ginkgo20 commented 11 months ago

sorry, i see no relation with my "4 Knights Position".

"(2 Knights vs 2 Bishops) Position" is no different from "(4 Knights) Position" because these positions rarely occur in games.