Page 2 of 2

Re: The optimalized Fat Fritz

Posted: Sun Dec 01, 2019 12:03 pm
by shrapnel
corres wrote:
Fri Nov 29, 2019 11:52 pm
In the cited article of Albert Silver you can see parameters of Fat Fritz on pictures.
Try them.
I did. They don't suit lc0 at all !
So that's a bust.
Still looking for best Settings for twin 2080 Tis Setups...

Re: The optimalized Fat Fritz

Posted: Tue Dec 03, 2019 7:57 am
by corres
shrapnel wrote:
Sun Dec 01, 2019 12:03 pm
corres wrote:
Fri Nov 29, 2019 11:52 pm
In the cited article of Albert Silver you can see parameters of Fat Fritz on pictures.
Try them.
I did. They don't suit lc0 at all !
So that's a bust.
Still looking for best Settings for twin 2080 Tis Setups...
Thanks for your information

Re: The optimalized Fat Fritz

Posted: Tue Dec 03, 2019 12:29 pm
by Albert Silver
shrapnel wrote:
Sun Dec 01, 2019 12:03 pm
corres wrote:
Fri Nov 29, 2019 11:52 pm
In the cited article of Albert Silver you can see parameters of Fat Fritz on pictures.
Try them.
I did. They don't suit lc0 at all !
So that's a bust.
Still looking for best Settings for twin 2080 Tis Setups...
I warned people to not use them for Leela, so I am not in the least surprised. As to the parameters, although they are the result of a very long CLOP tuning (170 hours), it is possible that one of the of the assumptions is simply weaker. My very limited resources sometimes makes testing the way I would like completely impossible. I chose FPU Strategy as Absolute, using also the predetermined value of -1 as selected by Deep Mind, and then tuned the other values around it. It is possible this is just weakening. I do have tests that show FF stronger than Test40 quite clearly, but the catch is that I tested in default lc0 settings for both. I will revisit my settings and a longer tune with no assumptions and see what comes. As to the result I mentioned, just for transparency:

Score of FF v281 vs lc0-v22 Leela 42850: 248 - 185 - 567 [0.531]
Elo difference: 21.92 +/- 14.15

1000 of 1000 games finished.


This was done at a fixed 1600 nodes per move, and was done on a slightly older build I should note. Anyhow, live and learn.

Re: The optimalized Fat Fritz

Posted: Tue Dec 03, 2019 8:59 pm
by corres
Albert Silver wrote:
Tue Dec 03, 2019 12:29 pm
...
I do have tests that show FF stronger than Test40 quite clearly, but the catch is that I tested in default lc0 settings for both. I will revisit my settings and a longer tune with no assumptions and see what comes. As to the result I mentioned, just for transparency:
Score of FF v281 vs lc0-v22 Leela 42850: 248 - 185 - 567 [0.531]
Elo difference: 21.92 +/- 14.15
1000 of 1000 games finished.

This was done at a fixed 1600 nodes per move, and was done on a slightly older build I should note. Anyhow, live and learn.
I think a test with fixed nodes can not replace the test with limited time.
Because in practice the games are played on time and not on nodes.

Re: The optimalized Fat Fritz

Posted: Wed Dec 04, 2019 3:50 am
by shrapnel
corres wrote:
Tue Dec 03, 2019 8:59 pm
I think a test with fixed nodes can not replace the test with limited time.
Because in practice the games are played on time and not on nodes.
+1.