TalkChess.com

v0.25.0-rc1

Training now outputs v5 format data. This passes the moves left information back to training. This also includes support for multiple sub formats,
including the existing standard, a new variant which can encode FRC960 castling, and also a further extension of that which tries to make training
data cannonical, so there aren't multiple positions that are trivially equivalent with different network inputs.

And the engine now has a UCI parameter UCI_chess960

As a matter of fact, we already trained a Chess960 network to test the implementation.
Take a look at our list of networks and download for example the network 712574.
It supports both Chess960 and the new moves-left-head.

If you want to see the moves-left-head in action, try these additional settings: With these settings, it will favor the quickest win whenever possible and you will most likely see much less shuffling in won positions.
Note, that this is a very small test run very early in its development. Try it against weaker opponents like Pedone etc.

edit: As a bonus, the GUI Nibbler now also supports Chess960

Thanks. Allie 0.5 did amazingly well for me just using the standard 42850 net, but I reckon there is +100 to 150 Elo to be gained on a chess960 trained net.

Kiudee wrote: ↑Sat Apr 18, 2020 1:06 pm As a matter of fact, we already trained a Chess960 network to test the implementation.
Take a look at our list of networks and download for example the network 712574.
It supports both Chess960 and the new moves-left-head.

Yes I tried that net at chess960 and it wasn't great.

I'm currently running it against 591226 to establish a baseline, and we'll see how it improves from there.

Both MiniHuman and leela2200 on lichess play chess960 on lc0 0.24, so the capability has already been there.

dkappe wrote: ↑Mon Apr 20, 2020 8:25 pm Both MiniHuman and leela2200 on lichess play chess960 on lc0 0.24, so the capability has already been there.

Amazing, I did not see that anywhere.

Kiudee wrote: ↑Sat Apr 18, 2020 1:06 pm As a matter of fact, we already trained a Chess960 network to test the implementation.
Take a look at our list of networks and download for example the network 712574.
It supports both Chess960 and the new moves-left-head.

<snip>
Note, that this is a very small test run very early in its development. Try it against weaker opponents like Pedone etc.

Yes I did try 712574 and it was pretty weak compared to what Lc0 is capable of in normal chess. But the first results are up with 591226

http://ccrl.chessdom.com/ccrl/404FRC/

More still to come.

If there is ever a trained 960 network it will be interesting to see the improvement. The effect would be earlier in the games; once they get to a certain point they are mostly indistinguishable from standard games .

The problem is that your comparisons is wrong.

But feel free to show me which of your different networks was stronger, than 712574 or 712576, after only the first day of training.

712576 looks stronger than 712576.

Not wrong at all, I don't have all day and banks of machines to test every network. I ran just 712574 as suggested above by Kiudee, and 591226.

591226 will act as a baseline, and once an extensively trained network is available for chess960 we'll see what improvement there is.

Good that chess 960 improve tactics.
And LC0 would avoid things like this:

[White "Stockfish 150420 64 POPCNT"] 3 cores
[Black "LC0 30x384"] (3350) 2x RTX 2080 Ti
[Result "1/2-1/2"]
[ECO "B14"]
[TimeControl "60+1"]
No tb
Game started after 1.e3

[pgn]1. e3 Nf6 2. Nf3 c5 3. c4 g6 4. d4 cxd4 5. exd4 d5 6. cxd5 Nxd5 7. Qb3 Bg7 8.
Bc4 e6 9. Bg5 Qb6 10. Nc3 h6 11. Bd2 Ne7 12. Qxb6 axb6 13. Nb5 O-O 14. Nc7 Ra7
15. Nb5 Ra8 16. Nc7 Ra7 17. Nb5 Ra8 1/2-1/2[/pgn]