Minic version 3

[xr_a_y]

Recently, Minic 3.24 has been released for FRC event at TCEC.

It appears this version is finally quite near the first "MinicNNUE 2.53" using SF implementation using a Sergio V net.

Code: Select all

Rank Name                               Elo     +/-   Games   Score    Draw 
   0 minic_3.24                          
   1 minicNNUE_2.53_nn-97f742aaefcd      30      11    1670   54.3%   55.1% 
   2 minicNNUE_2.53_napping_nexus       -75      12    1670   39.4%   49.9% 
   3 minicNNUE_2.53_nascent_nutrient   -151      14    1670   29.5%   34.9% 

But the gap is not closed yet ! it took me 2 years ... from august 2020 to probably august 2022 just to reach my first "copy/paste" NNUE impl ... So it was not a meaningless decision to detach from SF impl ...

[xr_a_y]

I'm happy to announce Minic 3.27 using new net named "natural naughtiness".
This version shall be quite stronger than 3.22 or 3.24, maybe +40 to +60Elo at STC.

The net is using the same architecture as previous one, with a common input layer but various inner layers depending on number of pieces on the board (here with only 2 buckets). Only last bucket was trained with lambda!=1, the first one is always only trained with lambda=1. Data used here were a mix of former data (used for "nylon nonchalance") with additionnal 2B FRC fixed nodes using previous net.

This version also comes with some good search patches and minor fixes. I'm also preparing for some experiment : persistent hash table, game history knowledge (boom / moob), ...

I think next update will take some time as I also want to experiment for a while with fixed nodes data and with lambda < 0.5 for my last subnet based on Minic testing process pgn (much deeper depth than usually generated data).

Executables here : https://github.com/tryingsomestuff/Mini ... ter/Minic3

Have fun !

[Dariusz]

I'm happy to announce Minic 3.27 using new net named "natural naughtiness".
This version shall be quite stronger than 3.22 or 3.24, maybe +40 to +60Elo at STC.

The net is using the same architecture as previous one, with a common input layer but various inner layers depending on number of pieces on the board (here with only 2 buckets). Only last bucket was trained with lambda!=1, the first one is always only trained with lambda=1. Data used here were a mix of former data (used for "nylon nonchalance") with additionnal 2B FRC fixed nodes using previous net.

This version also comes with some good search patches and minor fixes. I'm also preparing for some experiment : persistent hash table, game history knowledge (boom / moob), ...

I think next update will take some time as I also want to experiment for a while with fixed nodes data and with lambda < 0.5 for my last subnet based on Minic testing process pgn (much deeper depth than usually generated data).

Executables here : https://github.com/tryingsomestuff/Mini ... ter/Minic3

Have fun !

Thank you, Vivien
I'll try compile for Mac and test it!

[matejst]

Vivien,

First, I would like to congratulate you: Minic has become a mature, featured and and well rounded engine.

Can you tell me what are the differences between Minic 3.28 and 3.27? Have you changed the net? I saw Minic was a bit faster, but I also noticed some differences in evaluation, I guess they are the consequences of a somewhat different search but I would like to be sure.

Best wishes.

[xr_a_y]

Vivien,

First, I would like to congratulate you: Minic has become a mature, featured and and well rounded engine.

Can you tell me what are the differences between Minic 3.28 and 3.27? Have you changed the net? I saw Minic was a bit faster, but I also noticed some differences in evaluation, I guess they are the consequences of a somewhat different search but I would like to be sure.

Best wishes.

Thanks !

Indeed I'm making some search patches currently, more dev version will come before end of summer (no net update currently).
For testing, i'd suggest to stick to the official release 3.27.

Regards

[Frank Quisinsky]

*** Average game length: All: 88 moves (wins: 97 moves / draws: 88 moves / losses: 73 moves)

Draws = 88 moves !!!
Clearly improved, not sure what you do but for me it seems Contempt = 0 works now.

Stats comes from (not ready for Minic 3.27):
https://www.amateurschach.de/fling/fcp_ ... ecret.html

Most are attackers!

The same tourney I will replay with a test-set.
For that reason move average is important ... so I can better find out bad test-set positions (draw quote to high or move average to hight). Important is to have stronger engines like to play more defense chess in the group of engines (Minic or Seer, should be enough). Need a good mix from all!

Also great for Minic:
The quantity of fast lost games is lower from Minic 3.27 to other Minic versions.

Now, Frank is happy with Minic!
One engine more I can used for find out test-set positions.

With other words ...
Thanks Vivien that you thinking on it.

I like Minic really a lot because the very late mid-games are fantastic.
Congratulation for Elo gain (your latest versions).

Best
Frank

PS: In the last months I made some opening analyzes. I found out that around 10% of interesting openings (full board, aggressiveness and complicated lines) ended without contempt to fast. Only 10% not more.

Can be to 85% avoided if such a rule is in sources:
If min. 6 pawns and queen on the board avoid 3-fold!

Because in 85% interesting and aggressive openings are 6 pawns and queen (both sites) on the board.
Such a rule can be interesting for attackers, playing with Contempt = 0.

[xr_a_y]

Hi all,

a minor official update of Minic is available : Minic 3.30
Some search patches, fixes, and some speed up but I don't expect more than +10Elo.

Executables here : https://github.com/tryingsomestuff/Mini ... ter/Minic3

[matejst]

Thanks, Vivien. I really enjoy using Minic now. Vous avez fait un bon bout de chemin ces dernières années.

[Werner]

Thanks Vivien, looks too strong for league 2!

[Damir]

Hi all,

a minor official update of Minic is available : Minic 3.30
Some search patches, fixes, and some speed up but I don't expect more than +10Elo.

Executables here : https://github.com/tryingsomestuff/Mini ... ter/Minic3

Vivien your Minic github website is a mess.. Why not post the most recent and latest version on your website instead of 1000 others as well?

Anyway thanks for the new engine and keep up the good work.