Incorrect calculation of the next move

[alexiwonder]

Hi Gavin,
Thanks so much for developing this tiny but powerful engine.

I have been enjoying using it, however, today I noticed that the engine calculates the next move of the below position very much incorrectly.
Position: 2k4r/2p1bpp1/p7/1p2n2b/3pPP2/P2P3q/1PP4N/R3QRK1 b - f3 0 24

I checked it with the two engines below and they didn't suggest that move even with the lowest rating (1300):

• https://nextchessmove.com
• https://www.365chess.com/play_computer_online.php

I also maxed "thinking time", but got the same result.
I am not sure if you answer this type of question, but thought I ask anyway since it looks like a bug.

Appreciated!

[tissatussa]

I also maxed "thinking time", but got the same result.

i didn't .. i guess you used the BetaFish version at https://betafish.gavinong.com .. it gives Ng6 at 1 second thinking time, but Qxh2+ at 3,5,10 sec .. could it be different for you because our hardware / connection differs ? I doubt it .. you didn't mention your BetaFish moves .. i just did an analyses with the new and very strong Shashchess GZ NNUE engine, for reference :

indeed, many other moves are better, they even lead to a quick mate, but both Ng6 (11-th) and Qh2+ (12-th) appear in this top-12 list, being clearly winning also .. it seems this position is much better for Black in many variations .. so, should BetaFish choice really be considered a bug ?

[alexiwonder]

To see where Qh2 could lead, I had betafish to play against itself. It turned out that it would lead to checkmate after notably long trails of moves.
So, fair enough; it is not a bug.
However, I still feel it can be improved.
@Strryke Any chance you can have a look?

[Strryke]

Hey, thanks for this, it's nice to know people are playing around with Betafish.

I went to take a quick check, basically, it's not a bug — the engine just doesn't look far enough to see the mate in 8 at a lower thinking time and thus doesn't find a better move than Ng6.

I maxed out the thinking time and the engine finds the best move at Depth 7, while at a thinking time of 1 second, the engine only manages to reach depth 4 and decides Ng6 is the best move. I've attached the logs below.

Thanks for the question! I always love revisiting this project.