Stockfish uses the NNUE evaluation when the position is balanced, then move on classical evaluation when the position is unbalanced.
Why does this happen? Can we train a particularly effective evaluation net for unbalanced positions? A net with intrinsic contempt!
Can we train a NNUE net for playing unbalanced game?
Moderators: hgm, Rebel, chrisw
-
- Posts: 85
- Joined: Mon May 13, 2019 6:05 pm
- Full name: Marco Giorgio
-
- Posts: 3186
- Joined: Sat Feb 16, 2008 7:38 am
- Full name: Peter Martan
Re: Can we train a NNUE net for playing unbalanced game?
You haven't read about dkappes nets DarkHorse 0.1, 0.2, NightNurse 0.1, 0.2, have you?Marcus9 wrote: ↑Thu Sep 17, 2020 8:35 pm Stockfish uses the NNUE evaluation when the position is balanced, then move on classical evaluation when the position is unbalanced.
Why does this happen? Can we train a particularly effective evaluation net for unbalanced positions? A net with intrinsic contempt!
Maybe he's reading this and is willing to tell a little once in a while again about his work, I like all of his nets very much. But with hybrid eval like SF's you won't profit from nets for unbalanced positions, cause with hybrid NNUE eval and therefore use of nets is active only in balanced positions.
Which I still don't like for analysis as the only one option, but there are many branches of SF's letting the user choose between hybrid, classical and pure NNUE eval,e.g. Ronald de Man's Cfisch, another one branch in this regard is dkappes NetRunner, with which you can adopt the tresholds for using classical or NNUE eval manually by UCI option.
Peter.
-
- Posts: 291
- Joined: Wed May 08, 2013 6:49 am
Re: Can we train a NNUE net for playing unbalanced game?
I assume that what you mean by unbalanced is winning or losing positions. Yes, it makes sense to train NNUE particularly for losing positions because it can prove useful for handicaps. Normal AB evals handling of handicap is not optimal because of their tendency to take the bait of exchanging pieces for little material improvement at the expense of simplifying the endgame for the humans advantage.Marcus9 wrote: ↑Thu Sep 17, 2020 8:35 pm Stockfish uses the NNUE evaluation when the position is balanced, then move on classical evaluation when the position is unbalanced.
Why does this happen? Can we train a particularly effective evaluation net for unbalanced positions? A net with intrinsic contempt!
-
- Posts: 85
- Joined: Mon May 13, 2019 6:05 pm
- Full name: Marco Giorgio
Re: Can we train a NNUE net for playing unbalanced game?
How can we train this NNUE?Jhoravi wrote: ↑Fri Sep 18, 2020 6:16 amI assume that what you mean by unbalanced is winning or losing positions. Yes, it makes sense to train NNUE particularly for losing positions because it can prove useful for handicaps. Normal AB evals handling of handicap is not optimal because of their tendency to take the bait of exchanging pieces for little material improvement at the expense of simplifying the endgame for the humans advantage.Marcus9 wrote: ↑Thu Sep 17, 2020 8:35 pm Stockfish uses the NNUE evaluation when the position is balanced, then move on classical evaluation when the position is unbalanced.
Why does this happen? Can we train a particularly effective evaluation net for unbalanced positions? A net with intrinsic contempt!
-
- Posts: 5563
- Joined: Tue Feb 28, 2012 11:56 pm
Re: Can we train a NNUE net for playing unbalanced game?
It's just a speed optimisation that happens to gain Elo. It doesn't have much if anything to do with how networks are trained.