Marvin 6.3.0

[Martin]

A new version of Marvin is now available on GitHub. According to my test gauntlet it is about 30 elo stronger than the previous version.

Changes:
* New NNUE architecture, (768->1024)x2->1
* Switch back to pure NNUE evaluation
* Remove scaling of NNUE scores
* Remove delta pruning
* Include checks in first iteration of quiscence search
* Added possibility to generate and rescore training data
* Fixed compile problem with Android NDK
* Fixed a bug when evaluating piece-square tables
* Fixed a FEN parsing bug

Binaries: https://github.com/bmdanielsson/marvin- ... tag/v6.3.0

[Gabor Szots]

Hi Martin,

The new executables are much smaller than those of v 6.2.0. Also, v 6.3.0 plays much worse than expected. Obviously I have to download a network as well but I can't find it.

[Martin]

Hi Gabor,

The network is embedded in the binary, so you don't have to download anything extra. The reason the executable is so much smaller is because the new networks are only around 1.5 MB in size (compared to 20 MB before).

Which executable are you using, modern or avx2? Linux or Windows?

[Gabor Szots]

I use the avx2 exe under Windows. GUI is Arena and I use Marvin in xboard mode with this ini file:

HASH_SIZE=256
LOG_LEVEL=0
SYZYGY_PATH=F:\Sakk\syzygy
NUM_THREADS=1

[Gabor Szots]

Food for thought:

In a 720-game tornament, Marvin played the first 120 games in WB mode, scoring 46 points. Then I switched to UCI mode and in the remaining 600 games it scored 271 points. 38,3 % vs. 45,2 %. Of course a statistical fluke cannot be excluded.

[Martin]

Thank you for the info! My initial suspicion was actually that there were something wrong with my Xboard implementation since I do all my testing using UCI. However, so far Xboard performance is similar to UCI performance in my testing.

[Gabor Szots]

My test have confirmed your results. I played two 100-game matches using Marvin UCI and Marvin WB. Here are the results:

Under Arena (level: 90s + 1s, book: LowDraw1000 by Chris Wittington)

Engine Score Ma
1: Marvin 6.3.0 64-bit UCI 52,0/100 ·······
2: Marvin 6.3.0 64-bit WB 48,0/100 8-12-80

100 games played / Tournament is finished
Name of the tournament: Marvin UCI vs Marvin WB under Arena
Level: Blitz 1:30/1

Under Cute Chess (level: 30s + 0.2s, book: LowDraw1000 by Chris Wittington)

Score of Marvin_6.3.0-x64 UCI vs Marvin_6.3.0-x64 WB: 17 - 11 - 72 [0.530]
... Marvin_6.3.0-x64 UCI playing White: 9 - 4 - 37 [0.550] 50
... Marvin_6.3.0-x64 UCI playing Black: 8 - 7 - 35 [0.510] 50
... White vs Black: 16 - 12 - 72 [0.520] 100
Elo difference: 20.9 +/- 36.0, LOS: 87.2 %, DrawRatio: 72.0 %
SPRT: llr 0 (0.0%), lbound -inf, ubound inf
100 of 100 games finished.

So I don't think you have to worry about your WB implementation being broken.

[chessica]

Hi, I was able to find and fix a bug in the sources, I'm really happy.

Now it works!

[d]5k2/pp3N2/5K2/5B2/8/8/8/8 w - - 0 29
#9


Marvin 6.3.0:

8/8 00:00,011 3k 304k +7,40 Sf7-d6 a7-a5 Sd6xb7 a5-a4 Lf5-e6 Kf8-e8 Sb7-d6+ Ke8-d8
21/34 00:15,250 6.107k 400k +10,60 Sf7-d6 a7-a5 Lf5-g6 a5-a4 Sd6-b5 Kf8-g8 Sb5-a3 b7-b5 Lg6-f7+ Kg8-h8 Sa3xb5 Kh8-h7 Sb5-d6 a4-a3 Sd6-b5 a3-a2 Lf7xa2 Kh7-h6 La2-f7 Kh6-h7 Sb5-d6 Kh7-h6 Sd6-f5+ Kh6-h7 Lf7-b3 Kh7-h8
21/32 00:25,859 10.330k 399k +10,65 Lf5-d7 b7-b5 Ld7xb5 a7-a6 Lb5-d7 a6-a5 Sf7-g5 a5-a4 Ld7xa4 Kf8-g8 Sg5-f7 Kg8-f8 La4-d7 Kf8-g8 Sf7-e5 Kg8-f8 Ld7-g4 Kf8-g8 Lg4-e6+ Kg8-h8
30/46 03:07,561 70.823k 377k +199,81 Lf5-d7 b7-b5 Sf7-g5 b5-b4 Sg5-h7+ Kf8-g8 Kf6-g6 Kg8-h8 Ld7-e6 b4-b3 Le6xb3 a7-a5 Sh7-f8 a5-a4 Lb3-c4 a4-a3 Kg6-h6 a3-a2 Sf8-g6+