Page 11 of 16

Re: Minic version 2

Posted: Mon Sep 28, 2020 6:22 pm
by xr_a_y
Werner wrote: Mon Sep 28, 2020 6:11 pm Hi Vivien,
thanks for the new version.
How do I see that the net is correct loaded?
e.g. in command promt:
info string This is Minic version 2.50 (NNUE available)
setoption name NNUEFile value nn.bin
ucinewgame
go depth 1
info depth 1 score cp 1 time 1 nodes 23 nps 22999 seldepth 1 pv d2d4 tbhits 0
bestmove d2d4
Very good question. Thanks !

Minic will display
"This is Minic version 2.50 (NNUE available)"
as soon as the binary is compiled with NNUE activated and this is the case for all the ones I give.

To double check a net is loaded I recommend this two tricks :
- be sure Minic is slower than usual (can be done without changing anything, just looking at nps).
- you can try to use the -quiet 0 command line option (just for a test, not always activated). You will then see

Code: Select all

info string Info  2020-09-28 18:21:31-663: Locking other threads
info string Info  2020-09-28 18:21:31-663: Calling main thread search
info string Info  2020-09-28 18:21:31-663: Search launched for thread 0
info string Info  2020-09-28 18:21:31-663: NNUE evaluation using Tourney/nn.bin enabled.    <============== HERE !
info string Info  2020-09-28 18:21:31-663: Search params :
info string Info  2020-09-28 18:21:31-663: requested time  2592000000
info string Info  2020-09-28 18:21:31-663: requested depth 50
Do you think it might be helpful in next release to always display such an information ?

Re: Minic version 2

Posted: Mon Sep 28, 2020 8:18 pm
by Werner
Yes :wink:

Re: Minic version 2

Posted: Fri Oct 02, 2020 7:42 pm
by CMCanavessi
xr_a_y wrote: Mon Sep 28, 2020 6:22 pm Do you think it might be helpful in next release to always display such an information ?
Absolutely!

Do you have plans to make your own network in the future? How strong is plain 2.50 without net?

Re: Minic version 2

Posted: Fri Oct 02, 2020 8:45 pm
by xr_a_y
CMCanavessi wrote: Fri Oct 02, 2020 7:42 pm
xr_a_y wrote: Mon Sep 28, 2020 6:22 pm Do you think it might be helpful in next release to always display such an information ?
Absolutely!

Do you have plans to make your own network in the future? How strong is plain 2.50 without net?
I did 2 nets by myself already (see https://github.com/tryingsomestuff/NNUE-Nets). One from SF data, using SF learner and another from Minic data with merged learner inside Minic (so "pure" Minic). They both perform well under SV nets of course !

Code: Select all

Rank Name                                 Elo     +/-   Games   Score    Draw 
   1 minic_2.50_nnue_nn-97f742aaefcd      150      41     201   70.4%   32.3% 
   2 minic_2.50_nnue_napping_nexus         42      37     201   56.0%   40.3% 
   3 minic_2.50_nnue_nascent_nutrient     -19      39     200   47.3%   35.5% 
   4 minic_2.50                          -177      45     202   26.5%   26.2% 
2.50 with its standard eval is maybe +10 or +15 versus 2.48. At blitz TC is should be nearly as good as Texel.

Re: Minic version 2

Posted: Fri Oct 02, 2020 10:30 pm
by CMCanavessi
xr_a_y wrote: Fri Oct 02, 2020 8:45 pm
CMCanavessi wrote: Fri Oct 02, 2020 7:42 pm
xr_a_y wrote: Mon Sep 28, 2020 6:22 pm Do you think it might be helpful in next release to always display such an information ?
Absolutely!

Do you have plans to make your own network in the future? How strong is plain 2.50 without net?
I did 2 nets by myself already (see https://github.com/tryingsomestuff/NNUE-Nets). One from SF data, using SF learner and another from Minic data with merged learner inside Minic (so "pure" Minic). They both perform well under SV nets of course !

Code: Select all

Rank Name                                 Elo     +/-   Games   Score    Draw 
   1 minic_2.50_nnue_nn-97f742aaefcd      150      41     201   70.4%   32.3% 
   2 minic_2.50_nnue_napping_nexus         42      37     201   56.0%   40.3% 
   3 minic_2.50_nnue_nascent_nutrient     -19      39     200   47.3%   35.5% 
   4 minic_2.50                          -177      45     202   26.5%   26.2% 
2.50 with its standard eval is maybe +10 or +15 versus 2.48. At blitz TC is should be nearly as good as Texel.
Awesome! And when (if ever) do you plan to make the switch (pack minic with its own network and make it use the NN as default)?

Re: Minic version 2

Posted: Fri Oct 02, 2020 10:34 pm
by Tony P.
Fingers crossed for the RL improvement of Nascent Nutrient :)

Re: Minic version 2

Posted: Sat Oct 03, 2020 8:29 am
by xr_a_y
CMCanavessi wrote: Fri Oct 02, 2020 10:30 pm
xr_a_y wrote: Fri Oct 02, 2020 8:45 pm
CMCanavessi wrote: Fri Oct 02, 2020 7:42 pm
xr_a_y wrote: Mon Sep 28, 2020 6:22 pm Do you think it might be helpful in next release to always display such an information ?
Absolutely!

Do you have plans to make your own network in the future? How strong is plain 2.50 without net?
I did 2 nets by myself already (see https://github.com/tryingsomestuff/NNUE-Nets). One from SF data, using SF learner and another from Minic data with merged learner inside Minic (so "pure" Minic). They both perform well under SV nets of course !

Code: Select all

Rank Name                                 Elo     +/-   Games   Score    Draw 
   1 minic_2.50_nnue_nn-97f742aaefcd      150      41     201   70.4%   32.3% 
   2 minic_2.50_nnue_napping_nexus         42      37     201   56.0%   40.3% 
   3 minic_2.50_nnue_nascent_nutrient     -19      39     200   47.3%   35.5% 
   4 minic_2.50                          -177      45     202   26.5%   26.2% 
2.50 with its standard eval is maybe +10 or +15 versus 2.48. At blitz TC is should be nearly as good as Texel.
Awesome! And when (if ever) do you plan to make the switch (pack minic with its own network and make it use the NN as default)?
For now this is not something I want. Please see http://talkchess.com/forum3/viewtopic.p ... 67#p863067

Re: Minic version 2

Posted: Wed Oct 07, 2020 1:30 am
by OliverBr
xr_a_y wrote: Tue Mar 31, 2020 4:45 pm So here we are ! Minic version 2 is available :D . The single file minic.cc has been split into many source files, thus increasing maintainability.
This version is functionally the same as 1.55 (including multiPV development).
Thank you very much to your strong engine, which is not another SF clone.
It is also, of course, time to say good bye to the good old "A complete 2000 lines of code engine" thread.
Code has is now over 10k lines. Would you still denote your engine as "minimalistic" chess engine"? Have the goals changed?
Which was the last most "minimalistic" version? I would really like to study it.

Re: Minic version 2

Posted: Wed Oct 07, 2020 7:55 am
by xr_a_y
OliverBr wrote: Wed Oct 07, 2020 1:30 am
It is also, of course, time to say good bye to the good old "A complete 2000 lines of code engine" thread.
Code has is now over 10k lines. Would you still denote your engine as "minimalistic" chess engine"? Have the goals changed?
Which was the last most "minimalistic" version? I would really like to study it.
We can say the goal is not to be "minimal" anymore since version 2.
I'm not sure this kind of minimalism did not allow Minic to go stronger, but for sure a good old big refactoring raised some bug and ideas !

It you are looking for strong small code base engines, I recommend Xiphos or Demolito or Weiss as a starting point.
If you are interested in very small engines, nanochess and micromax are the place to go.

In the middle, there are (quite a lot of) medium strength engines with around 2000 or 3000 lines of codes, like Minic at some point.

Re: Minic version 2

Posted: Wed Oct 07, 2020 1:48 pm
by OliverBr
xr_a_y wrote: Wed Oct 07, 2020 7:55 am It you are looking for strong small code base engines, I recommend Xiphos or Demolito or Weiss as a starting point.
Weiss ist just great and amazingly small and strong. For Xiphos I read rumors about being a clone in another forum, even though it's open source, which makes the situation even more difficult.

Furthermore, they say on http://rebel13.nl/misc/sim2019.html.

"Note the extreme high similarity of 83.57% between Schooner and Xiphos 0.3"

83.57% is quite a number. I don't know what to think about it.

Note that OliThink doesn't have more than 30% similarity with any other engine.