Minic version 3

[Guenther]

I understand and thank you for your xr_a_y. Version it Cute Chess 1.2.0 Using Qt version 5.15.0 not Cute Chess 1.2.0 Using Qt version 5.14.2.

What Vivien tried to explain to you for a while is that you should best use a full path for being on the safe side
(also you have to check 'forceNNUE' of course - you did not in your first screenshot)


Then you'll also notice (like Vivien said) that the 'speed' of the program will be different (slower!), if the nnue is used
(right-click 'show image' gives full size image - here speed of nnue is around 50% of normal evaluation)


CuteChessGUI does not save full debugs (only cli does), or acknowledge the net is loaded.
Of course you could still feed the necessary (and same as to Cute) commands via cmd and you'll see that Minic tells that it has loaded the net (or not).

[Guenther]

...
(also you have to check 'forceNNUE' of course - you did not in your first screenshot)
...

I want to correct above, forceNNUE is not needed for running the NNUE. It decides about using (default) hybrid mode or pure NNUE.

from Minics github option explanations:

Code: Select all

-forceNNUE [0 or 1] (default is false): if a NNUEFile is loaded, forceNNUE equal true will results in a pure NNUE evaluation, while the default is hybrid evaluation

[Guenther]

Thank you very much Guenther. How to check with a command cmd.

see below: blue = my command to cmd
(not all commands are needed for all programs, but it is a good habit to know them and use them in that sequence)

I hope this is sufficient I don't think you can get more convenient support for this

1. Here I tell Minic to find the nnue file and you can see the effect in nps speed
---
info string This is Minic version 3.04 (NNUE available)
info string Info 2021-03-07 17:07:47-477: No NNUE net loaded, using standard evaluation
uci
id name Minic 3.04
id author Vivien Clauzon
option name Level type spin default 100 min 0 max 100
option name UCI_LimitStrength type check default false
option name UCI_Elo type spin default 1500 min 500 max 2800
option name Hash type spin default 128 min 1 max 256000
option name Threads type spin default 1 min 1 max 255
option name UCI_Chess960 type check default false
option name Ponder type check default false
option name MateFinder type check default false
option name MultiPV type spin default 1 min 1 max 4
option name RandomOpen type spin default 0 min 0 max 100
option name MinMoveOverHead type spin default 50 min 10 max 1000
option name Contempt type spin default 12 min -50 max 50
option name ContemptMG type spin default 12 min -50 max 50
option name SyzygyPath type string default
option name NNUEFile type string default
option name forceNNUE type check default false
option name GenFen type check default false
option name GenFenDepth type spin default 8 min 2 max 20
option name RandomPly type spin default 0 min 0 max 20
option name StyleAttack type spin default 50 min 0 max 100
option name StyleComplexity type spin default 50 min 0 max 100
option name StyleDevelopment type spin default 50 min 0 max 100
option name StyleMaterial type spin default 50 min 0 max 100
option name StyleMobility type spin default 50 min 0 max 100
option name StylePositional type spin default 50 min 0 max 100
option name StyleForwardness type spin default 50 min 0 max 100
uciok
ucinewgame
isready
readyok
setoption name NNUEFile value C:/Engines/UCI/Minic_304-64/noisy_notch.bin
info string Info 2021-03-07 17:09:02-837: Loading NNUE net C:/Engines/UCI/Minic_304-64/noisy_notch.bin
position startpos
go infinite
info depth 1 score cp 16 time 1 nodes 24 nps 23999 seldepth 1 tbhits 0 pv e2e4
...snip...
info depth 15 score cp 43 time 1720 nodes 205806 nps 119654 seldepth 24 tbhits 0 pv e2e4 c7c5 g1f3 e7e6 b1c3 b8c6 a2a3 g8f6 f1b5 c6d4

2. Here I don't tell Minic to find any nnue file and you can see the effect in nps speed
---
info string This is Minic version 3.04 (NNUE available)
info string Info 2021-03-07 17:10:14-747: No NNUE net loaded, using standard evaluation
uci
id name Minic 3.04
id author Vivien Clauzon
option name Level type spin default 100 min 0 max 100
option name UCI_LimitStrength type check default false
option name UCI_Elo type spin default 1500 min 500 max 2800
option name Hash type spin default 128 min 1 max 256000
option name Threads type spin default 1 min 1 max 255
option name UCI_Chess960 type check default false
option name Ponder type check default false
option name MateFinder type check default false
option name MultiPV type spin default 1 min 1 max 4
option name RandomOpen type spin default 0 min 0 max 100
option name MinMoveOverHead type spin default 50 min 10 max 1000
option name Contempt type spin default 12 min -50 max 50
option name ContemptMG type spin default 12 min -50 max 50
option name SyzygyPath type string default
option name NNUEFile type string default
option name forceNNUE type check default false
option name GenFen type check default false
option name GenFenDepth type spin default 8 min 2 max 20
option name RandomPly type spin default 0 min 0 max 20
option name StyleAttack type spin default 50 min 0 max 100
option name StyleComplexity type spin default 50 min 0 max 100
option name StyleDevelopment type spin default 50 min 0 max 100
option name StyleMaterial type spin default 50 min 0 max 100
option name StyleMobility type spin default 50 min 0 max 100
option name StylePositional type spin default 50 min 0 max 100
option name StyleForwardness type spin default 50 min 0 max 100
uciok
ucinewgame
isready
readyok
position startpos
go infinite
info depth 1 score cp 102 time 1 nodes 21 nps 20999 seldepth 1 tbhits 0 pv b1c3
...snip...
info depth 15 score cp 58 time 1740 nodes 512812 nps 294719 seldepth 28 tbhits 0 hashfull 41 pv g1f3 d7d5 d2d4 e7e6 e2e3 g8f6 f1d3 f8b4 c1d2 b8c6 e1g1 e8g8 c2c4 d5c4 d2b4

[Guenther]

I have something like this at once (like below) and farther I don't know what I am supposed to do.

well, you can start just by double clicking the Minic windows binary
(otherwise you would first have to direct to the final folder and binary and it will be loaded in the terminal)

[Guenther]

You did as you said - it doesn't work. Or it's an engine not on my processor.

But this is just the beginning - read my post again, you need to feed commands - I had it explained all.
By default it will not load any NNUE...

If you don't know how to feed commands just type them (I won't explain how to copy/paste them into the cmd window)
and hit return after each command, otherwise it will not be sent.

[Guenther]

I have this information - info string unrecognised comand setopion.

It is setoption (not setopion). Please don't post each of your typos now in this thread.
(I feel a bit tired now... like 'setopium')

[xr_a_y]

Thanks also to the author of the engine.

Thank you for using it

[xr_a_y]

And where the latest version of the program.

"3.06 alpha" was just a test for tcec, official release is coming soon with a new net also.

[xr_a_y]

Hi all,

I'm happy to announce here Minic 3.06 with a new net named "Nocturnal Nadir".

https://github.com/tryingsomestuff/Mini ... s/tag/3.06
https://github.com/tryingsomestuff/NNUE-Nets

This release introduces some bug fixes and some features.
The new net named "Nocturnal Nadir" is available here (see https://github.com/tryingsomestuff/NNUE-Nets) and plays at least +20Elo above "Noisy Notch".

Minic 3.06 can now play in Armageddon mode (using HCE or NNUE). This requieres a WDL model which was fitted using recent rating lists data. WDL outputs can be displayed using the "WDL_display" option (but they are not "true" WDL outputs of course). Maybe some fun tourney soon !

Minic 3.06 has also the possibility to use UCI_Opponent or UCI_RatingAdv (for TCEC S21). Using this, Minic will adapt contempt to its opponent.

Another added feature is the possibility to make Minic weaker without any randomness, only based on max nodes. To activate this, use the "nodesBasedLevel" option (with Level or UCI_LimitStrength and UCI_Elo options of course). A future work will be to better fit Minic level to FIDE rating.

Help / documentation of Minic has been updated quite a lot, I hope informations needed are available there.

What didn't work well since release 3.04 :
- train smaller nets (they are quite good at STC but fails at LTC), Minic 3.05 + "Nibbled Nutshel"l is such a try and is not officially released for this reason
- use multiple small nets depending on game phase (I tried 2 or 3). It seems a single net well trained ok already ok, no need to "help" the train process by separating game phase.
- search parameters tuning using some very good tools already advertised on this forum. I didn't get anything better than current parameters unfortunatly for now.

Work in progress for next release includes :
- better Elo fit for UCI_Elo option
- debug of "distributed" version (distributed memory version of Minic using MPI)
- data generation using noisy noise instead of HCE
- another try to quantization of the net

Many recent experiment shows that using something else than the "slylake" version (BMI2 + AVX2) to test NNUE leads to quite weaker results, I really encourage rating lists to only use the "slylake" executable when the hardware allows for it.

Have fun!

[Werner]

Thanks,
and the new net is hidden like an Easter Egg