Fat Fritz update

Discussion of anything and everything relating to chess playing software and machines.

Moderators: bob, hgm, Harvey Williamson

Forum rules
This textbox is used to restore diagrams posted with the [d] tag before the upgrade.
pohl4711
Posts: 1244
Joined: Sat Sep 03, 2011 5:25 am
Location: Berlin, Germany
Contact:

Re: Fat Fritz update

Post by pohl4711 » Thu Mar 05, 2020 3:47 pm

Albert Silver wrote:
Thu Mar 05, 2020 2:16 pm
Nordlandia wrote:
Thu Mar 05, 2020 10:35 am
bmp1974 wrote:
Thu Mar 05, 2020 10:02 am
The latest update is Fat Fritz 1.1 with changes to Cpuct, Fpu, Policy temp, NN cache and couple of others.
Tests are required to see where it stands.
Question is whether Kiudee settings work with Fat Fritz 🤔
Better than that: I ran Kiudee's tuner for 7 straight days on it to find its own 'Kiudee' settings.
Oh. Cool! I am curios to see, how much 1.1 is improved over 1.0. Parameter-tuning and a new net - very nice.

User avatar
Ozymandias
Posts: 1177
Joined: Sun Oct 25, 2009 12:30 am

Re: Fat Fritz update

Post by Ozymandias » Thu Mar 05, 2020 5:15 pm

I'll keep posted.

pohl4711
Posts: 1244
Joined: Sat Sep 03, 2011 5:25 am
Location: Berlin, Germany
Contact:

Re: Fat Fritz update

Post by pohl4711 » Thu Mar 05, 2020 6:35 pm

pohl4711 wrote:
Thu Mar 05, 2020 3:47 pm
Albert Silver wrote:
Thu Mar 05, 2020 2:16 pm
Nordlandia wrote:
Thu Mar 05, 2020 10:35 am
bmp1974 wrote:
Thu Mar 05, 2020 10:02 am
The latest update is Fat Fritz 1.1 with changes to Cpuct, Fpu, Policy temp, NN cache and couple of others.
Tests are required to see where it stands.
Question is whether Kiudee settings work with Fat Fritz 🤔
Better than that: I ran Kiudee's tuner for 7 straight days on it to find its own 'Kiudee' settings.
Oh. Cool! I am curios to see, how much 1.1 is improved over 1.0. Parameter-tuning and a new net - very nice.
Tuned parameters are (default in FatFritz 1.1 CUDA):
Fpu=0.74, CPuct=2.8 and PolicyTemperature=1.85.

More information here:
https://en.chessbase.com/post/fat-fritz ... t-fritz-jr

https://en.chessbase.com/portals/all/20 ... pdate4.png

Albert Silver
Posts: 2894
Joined: Wed Mar 08, 2006 8:57 pm
Location: Rio de Janeiro, Brazil

Re: Fat Fritz update

Post by Albert Silver » Thu Mar 05, 2020 6:43 pm

pohl4711 wrote:
Thu Mar 05, 2020 6:35 pm
pohl4711 wrote:
Thu Mar 05, 2020 3:47 pm
Albert Silver wrote:
Thu Mar 05, 2020 2:16 pm
Nordlandia wrote:
Thu Mar 05, 2020 10:35 am
bmp1974 wrote:
Thu Mar 05, 2020 10:02 am
The latest update is Fat Fritz 1.1 with changes to Cpuct, Fpu, Policy temp, NN cache and couple of others.
Tests are required to see where it stands.
Question is whether Kiudee settings work with Fat Fritz 🤔
Better than that: I ran Kiudee's tuner for 7 straight days on it to find its own 'Kiudee' settings.
Oh. Cool! I am curios to see, how much 1.1 is improved over 1.0. Parameter-tuning and a new net - very nice.
Tuned parameters are (default in FatFritz 1.1 CUDA):
Fpu=0.74, CPuct=2.8 and PolicyTemperature=1.85.

More information here:
https://en.chessbase.com/post/fat-fritz ... t-fritz-jr

https://en.chessbase.com/portals/all/20 ... pdate4.png
FPU Strategy is also set to Reduction. It is not in the image but is part of the tuning.
"Tactics are the bricks and sticks that make up a game, but positional play is the architectural blueprint."

cma6
Posts: 149
Joined: Thu May 29, 2014 3:58 pm

FF hardware requirements

Post by cma6 » Thu Mar 05, 2020 10:13 pm

To experienced FF users:
What is minimum hardware requirement (RAM and GPU card) required in order to expect reasonably good output from the Lc0 component of FF?

dkappe
Posts: 371
Joined: Tue Aug 21, 2018 5:52 pm
Full name: Dietrich Kappe

Re: Fat Fritz update

Post by dkappe » Fri Mar 06, 2020 1:51 am

I’ve been running some obscure lines in the Sicilian with recent t60 nets as black vs sf11. With a rtx2070 vs 8 cpu, t60 generally has an advantage, but in these Sicilian lines they haven’t done so well.

I decided to run sf11 vs both the latest Fat Fritz and a recent t60 net on v0.24 (both nets always playing black). The opening? The most common dozen variations for white at move 6 out of the Najdorf. TC 1 min + 1 sec.

It’s early yet. I plan on running 400 games for each net.

Code: Select all


   # PLAYER    :  RATING  ERROR  POINTS  PLAYED   (%)  CFS(%)    W    D    L  D(%)
   1 FFv471    :      24     57    54.5      96  56.8      62   20   69    7  71.9
   2 SF11      :       0    101    89.5     191  46.9      63   18  143   30  74.9
   3 62627     :     -27     57    47.0      95  49.5     ---   10   74   11  77.9

White advantage = -23.32 +/- 78.59
Draw rate (equal opponents) = 76.30 % +/- 3.08


carldaman
Posts: 1843
Joined: Sat Jun 02, 2012 12:13 am

Re: Fat Fritz update

Post by carldaman » Fri Mar 06, 2020 4:50 am


Tuned parameters are (default in FatFritz 1.1 CUDA):
Fpu=0.74, CPuct=2.8 and PolicyTemperature=1.85.
As far as I can see, the opencl default settings for FF 1.1 are:

Fpu=1.20, CPuct=3.0, PolicyTemp=2.2

Can anyone else verify this?
Thanks!

User avatar
AdminX
Posts: 5386
Joined: Mon Mar 13, 2006 1:34 pm
Location: Acworth, GA
Contact:

Re: Fat Fritz update

Post by AdminX » Fri Mar 06, 2020 11:03 am

I just played a LTC game with the latest Fat Fritz vs Komodo 13.3. I did not get a Sicilian but and interesting Ruy Lopez position was reached via an move order transposition. Fat Fritz played the black pieces.



In the position above Fat Fritz plays the Novelty 15. ... c6 here instead of Rb8.

Komodo followed up with 16. Nc4



Fat Fritz played the expected cxd5. Now Komodo surprised me with Nxd6 here, leading to an interesting exchange sac.



The game continued 17. ... Qxd6 18. Bxe5 Qd8 19. Bxh8



The game went on for 231 moves before ending in a draw.

"Good decisions come from experience, and experience comes from bad decisions."
__________________________________________________________________
Ted Summers

Hugo
Posts: 782
Joined: Tue Dec 01, 2009 10:10 am

Re: Fat Fritz update

Post by Hugo » Fri Mar 06, 2020 3:29 pm

Hi

same question I had. I started a test with FF1.1 on default settings. Results have not been promising (8 loss and 3 wins after 34 games), so I aborted the test, and started a new one with kiudee setting. I additionally prefer cpuct=1.900 from kiudee / Laskos setting which has shown a good performance on my tests.
https://docs.google.com/spreadsheets/d/ ... =208381013
Right now, FatFritz_1.1 plays way better after 48 games its 7 wins and 6 loss vs Stockfish 11.
conditions and a bit history on the spreadsheet.
C.K.
Nordlandia wrote:
Thu Mar 05, 2020 10:35 am
Question is whether Kiudee settings work with Fat Fritz 🤔

dkappe
Posts: 371
Joined: Tue Aug 21, 2018 5:52 pm
Full name: Dietrich Kappe

Re: Fat Fritz update

Post by dkappe » Fri Mar 06, 2020 3:43 pm

I’ve heard rumors that CB got the tuning wrong. Hopefully an update is coming soon.

Post Reply