I have just released version 5.2.0 of Marvin. Based on my tests it is about 70 elo stronger compared to Marvin 5.1.0.
The NNUE architecture have changed slightly so old nets won't work anymore. A new default net is avaibale that is trained on 1B depth 8 positions from Marvin 5.1.0 (using NNUE eval). As before the net is included in the binaries package, but it can now also be downloaded separately from here, https://github.com/bmdanielsson/marvin-nets .
I also fixed some very annoying bugs:
* A problem when allocating large hash tables which is causing Marvin to perform very poorly in the current TCEC Swiss event
* A problem that could cause random crashes on Windows
For those interested the training code for NNUE nets have also seen some improvements, most notably the addition of a factorizer.
Binaries: https://github.com/bmdanielsson/marvin- ... tag/v5.2.0
Marvin 5.2.0
Moderators: hgm, Rebel, chrisw
-
- Posts: 51
- Joined: Sun Jul 10, 2016 9:12 pm
- Location: Lund, Sweden
- Full name: Martin Danielsson
-
- Posts: 6808
- Joined: Wed Nov 18, 2009 7:16 pm
- Location: Gutweiler, Germany
- Full name: Frank Quisinsky
Re: Marvin 5.2.0
Hi Martin,
good news.
Engine Update added in FCP Tourney-2022.
Thank you and good luck!!
Best
Frank
good news.
Engine Update added in FCP Tourney-2022.
Thank you and good luck!!
Best
Frank
-
- Posts: 365
- Joined: Mon May 14, 2007 8:20 pm
- Full name: Boban Stanojević
Re: Marvin 5.2.0
Very interesting engine: a good balance of evaluation and speed.
Thanks!
Thanks!
-
- Posts: 6808
- Joined: Wed Nov 18, 2009 7:16 pm
- Location: Gutweiler, Germany
- Full name: Frank Quisinsky
Re: Marvin 5.2.0
Hi,
results looks great for me.
A very nice update from a very interesting engine!
Best
Frank
results looks great for me.
A very nice update from a very interesting engine!
Best
Frank
-
- Posts: 1365
- Joined: Sat Jul 21, 2018 7:43 am
- Location: Szentendre, Hungary
- Full name: Gabor Szots
-
- Posts: 7016
- Joined: Thu Aug 18, 2011 12:04 pm
- Full name: Ed Schröder
Re: Marvin 5.2.0
90% of coding is debugging, the other 10% is writing bugs.
-
- Posts: 7016
- Joined: Thu Aug 18, 2011 12:04 pm
- Full name: Ed Schröder
Re: Marvin 5.2.0
Estimated elo gain for Marvin 5.2 : +122
Wow...
Code: Select all
Gambit Rating List
Running : Gauntlet Marvin 5.2
Time Control : Time control : 40/120
Games : 1600
Results from file gauntlet-marvin.pgn:
No. Name Win Draw Loss Unf. Score Games %
----------------------------------------------------------
1 Marvin 5.2 +689 =627 -284 *0 1002.5 1600 62.7%
2 Seer 2.1.0 +49 =77 -74 *0 87.5 200 43.8%
3 Stash 31.0 +47 =72 -81 *0 83.0 200 41.5%
4 Zahak 8.0 +35 =91 -74 *0 80.5 200 40.2%
5 Weiss 1.4 +38 =82 -80 *0 79.0 200 39.5%
6 Wasp 4.50 +35 =83 -82 *0 76.5 200 38.2%
7 Halogen 10 +29 =83 -88 *0 70.5 200 35.2%
8 Berserk 4.2.0 +26 =70 -104 *0 61.0 200 30.5%
9 Clover 2.3.1 +25 =69 -106 *0 59.5 200 29.8%
Total Games: 1600
White Wins: 511 (31.9%)
Black Wins: 462 (28.9%)
Draws: 627 (39.2%)
Unfinished: 0 (0.0%)
Estimated elo gain for Marvin_5.2
Elo pool : 3146
Marvin 5.1 : 3105.0
Marvin_5.2 : 3227.5
Difference : 122.5
90% of coding is debugging, the other 10% is writing bugs.