Minic version 3

[xr_a_y]

Minic 3.17 1CPU and 4CPU gauntlets are running for the CCRL 40/15 list.

3.17 seems to perform well in div2, I'm glad you test it 4CPU on the 40/15 list thanks

[CMCanavessi]

3.13 and 3.14 are doing pretty good in my FRC list. Between Shredder 13 and Pedone 3.1, with 3.13 and 3.14 separated by 35 elo (3.14 on top). Waiting for 3.18 to upgrade...

[xr_a_y]

Minic 3.18 is released

As usual, binaries are there : https://github.com/tryingsomestuff/Mini ... ter/Minic3

This release is using a new (embedded) net named "nimble nothingness" (not compatible with previous release, because I'm back to clipped ReLu in order to retry quantization later...): this net was trained on 2B FRC Minic data (from "nucleated neurulation" at depth 10 to 14).

Minic 3.18 includes some juicy fast TC patches but I'm quite sure they don't scale at LTC unfortunatly.

It's anyway expected at least +70Elo versus previous release.

and I almost forgot, Happy new year!

[matejst]

Merci, Vivien ! Joyeux Nouvel An à toi également !

[Frank Quisinsky]

Merci, Vivien!
A bit to late for run-1 (my new rating-system) but will be used for run-2 start of February.
I am very happy that Minic 3.17 runs completly without any problems in my FCP Tourney-2022.
A very strong work you do for us!

Have a good new year!

Best
Frank

[xr_a_y]

One may ask where Minic is in terms of HCE versus NNUE evaluation after now one and a half year of training.
And the answer is "wow".

What are we comparing here ?:
- Minic 2.53 HCE : the standard evaluation of Minic from october 2020
- Minic 3.18 HCE : the current Minic HCE
- MinicNNUE 2.53 (SF NNUE impl copy/paste) using 3 different nets : one of the best SV net at that time, one home trained net based on SF data (napping nexus) and one home train net based on Minic HCE data at that time (nascent nutrient)
- Minic 3.18 with my last trained net (based on current Minic data)

Code: Select all

Rank Name                           Elo     +/-   Games   Score    Draw 
   1 minic_2.53_nn-97f742aaefcd     202      12    2581   76.2%   32.5% 
   2 minic_2.53_napping_nexus       113      11    2581   65.7%   37.0% 
   3 minic_3.18                      94      11    2581   63.3%   34.4% 
   4 minic_2.53_nascent_nutrient      1      11    2581   50.1%   33.0% 
   5 minic_2.53                    -214      13    2581   22.6%   26.2% 
   6 minic_3.18_HCE                -218      13    2581   22.2%   24.6% 

What do we learn ?:
- Minic HCE hasn't regress from 2.53 to 3.18, but hasn't progress either. This probably means that the various good search patches done this year does not work well with HCE. So clearly there is a search/eval interaction here where a better eval probably allows for a better (more agressive pruning / reduction?) search.
- The net trained for Minic 3.18 ("nimble nothingness" based on previous net data, itself based on previous net data, ...) now leads to an engine that is nearly as strong as MinicNNUE with a poorly trained SF data based net ("napping nexus").
- There is still 100Elo to be find, most notably in evaluation. This is the gap between current Minic with "nimble nothingness" and MinicNNUE with a good SF net (back in autumn 2020...)
- In the current state, Minic HCE is around 300Elo weaker than Minic using "nimble nothingness".

[Dokterchen]

What do we learn ?:
- Minic HCE hasn't regress from 2.53 to 3.18, but hasn't progress either. This probably means that the various good search patches done this year does not work well with HCE. So clearly there is a search/eval interaction here where a better eval probably allows for a better (more agressive pruning / reduction?) search.

Michael Borgstaedt writes in the readme file of his new NN version of his engine Little Goliath:

" The engine uses a neural network for positional evaluation. In direct
comparison with the old engine, which uses manually generated evaluation
functions, this engine achieves almost 100% in over 1000 test games (only
a few draw games). This is obviously significantly more than other
programmes have achieved when switching to neural networks. However, it
does not seem to be due to the net itself but to its effect on a pruning
system, which I call "path finding pruning" and which seems to depend
essentially on an exact position evaluation. If this pruning is
deactivated in both engines, the NN version "only" wins with about 70%.
If I leave the pruning activated in the classic engine and play against
the same version without it, the version with pruning is only just ahead
(approx. 52%). On fast hardware, this also corresponds to the results "

This is in line with your observation.

[xr_a_y]

I'm happy to finaly announced a new release !
So here is Minic 3.19.

A new net named "nylon nonchalance" is embedded in Minic 3.19. This net is trained on a mix of 4.2B DFRC data and 6B rescored T60T70wIsRight positions using Minic 3.18 (depth 5 to 14 depending on game phase). This net is using a new architecture with a common FT but various inner layers depending on number of pieces on the board (4 buckets). It is hard to compare with previous net as some juicy search patches as also been included but Minic 3.19 will play at least +30Elo versus Minic3.18 using "nimble nothingness" (I think around half of those is due to the new net). It is embedded in Minic 3.19.

The new net architecture is a bit slower so i'd expect :
- not that good results on non AVX2 hardware (beware of that for rating list)
- good scaling with threads number

The embedded net is available here : https://github.com/tryingsomestuff/NNUE-Nets
The executable are here : https://github.com/tryingsomestuff/Mini ... ter/Minic3

(I now pay for git LFS storage on github)

[Gabor Szots]

Congrats on the new version although it seems you could have stuffed 2 more in that size.

[xr_a_y]

Congrats on the new version although it seems you could have stuffed 2 more in that size.

Indeed, I'm trying big nets this time ... maybe I'll go back to smaller ones next...