Made some human-like, Maia-style NNUE nets (available on Lichess!)

Hi All!

Wanted to introduce myself by sharing something with the community. I made 3 NNUE nets for Cfish/Stockfish, all made from human games. The idea here is to allow the neural network to learn from <1200, <1600, and >2300 elo games and hopefully learn a human style from that.

I was initially a fan of the Maia nets; but i've found that it had endgame and tactical problems. These NNUE-style nets, with the help of stockfish search, should help address those limitations.

I would love for the community to give it a whirl and share some feedback!

If you want the nnue files, you can get them here: https://github.com/kennyfrc/nnue-networks/releases/tag/early-release. This is best enjoyed with the pure=yes and nnue=yes settings in Cfish.

One extra thing with the Lichess bots is that I've also added "dynamic nps" to my Cfish port, which means it has lower nps when time is low. Hopefully this simulates the blunders that might happen on low time controls. The average nps is about 200k.

- https://lichess.org/@/nnuehuman1200 (trained from 60m <1200 elo positions - lichess data)
- https://lichess.org/@/nnuehuman1500 (trained from 60m <1500 elo positions - lichess data)
- https://lichess.org/@/nnuehuman2400 (trained from 120m >2300 elo positions - kingbase data)

Hello Kenny, welcome to the forum. Tested the "kb2300-20201231-2057.bin" net:

Sf12 + your net vs SF12 default net

Score of sf12-kenny vs sf12-default:  Won=4 - Lost=109 - draws=29  [0.130]

Thanks for the test! Note that these networks are meant to be sparring partners, i.e. it will mimic the positional mistakes of that level. If you set Cfish at a nodestime of 50-100, it will also allow you to train your endgame conversion.

Here's quick test against shredder with elo of 1500.

cfish-2300 uses kb2300-20201230-0906.bin
cfish-1600 uses lichess1600-20201230-0908.bin
cfish-1200 uses lichess1200-20201230-0811.bin

Code:

Rank Name                          Elo     +/-   Games   Score    Draw
  1 cfish-2300                    338      69     180   87.5%   10.6%
  2 shredder-1500                  49      49     180   56.9%    8.3%
  3 cfish-1600                    -47      49     180   43.3%    7.8%
  4 cfish-1200                   -342      75     180   12.2%    5.6%

IMO chess programs that play like humans would be of value:

1. they'd be more fun to play against

2. they'd give more useful practice

At work, we sometimes play 10-minutes-each games in our lunch break. For me, the key is to play quickly and "reasonably accurately". I'd like to practice at this clock time, but if I play against a computer at this speed, I just get mauled.

I tried to play a 5 minute rated game against nnuehuman1500 on Lichess.org but I don't get a game?
I have tried Maia to, I liked it playing style but I don't like it that Maia never resigns. Even in hopeless lost positions.