Page 1 of 6
Stockfish NNUE style
Posted: Sat Aug 08, 2020 10:18 pm
by Rowen
What are the consequences of NNUE on the style of stockfish's play as opposed to the classical, or is it just a case of a stronger stockfish?
Re: Stockfish NNUE style
Posted: Sun Aug 09, 2020 1:06 am
by carldaman
NNUE seems stronger overall at all phases of the game, but especially in quiet positions where regular SF only sees 0.00, and transitioning to better endgames where it can grind out wins slowly but surely.
NNUE's superior strength also helps it discover hard to find moves much faster.
However, the current best nets have ways to go in achieving an aggressive style of play that creates chances vs weaker opposition. Its approach strikes me as solidly active, but with safety first in mind, until it finds a chink in the opponent's armor.
The night-nurse, Toga and lizard-fish nets that dkappe put together all look to be better at getting an edge against weaker opponents, especially with Black. This can be significant when it comes to analysis, when you're looking for ideas that have a lot of 'bite'.
The strongest nets may emphasize safety first a little too much considering how much stronger they are than the next best thing out there.
Re: Stockfish NNUE style
Posted: Sun Aug 09, 2020 10:14 am
by Rowen
carldaman wrote: ↑Sun Aug 09, 2020 1:06 am
NNUE seems stronger overall at all phases of the game, but especially in quiet positions where regular SF only sees 0.00, and transitioning to better endgames where it can grind out wins slowly but surely.
NNUE's superior strength also helps it discover hard to find moves much faster.
However, the current best nets have ways to go in achieving an aggressive style of play that creates chances vs weaker opposition. Its approach strikes me as solidly active, but with safety first in mind, until it finds a chink in the opponent's armor.
The night-nurse, Toga and lizard-fish nets that dkappe put together all look to be better at getting an edge against weaker opponents, especially with Black. This can be significant when it comes to analysis, when you're looking for ideas that have a lot of 'bite'.
The strongest nets may emphasize safety first a little too much considering how much stronger they are than the next best thing out there.
Interesting,thank you for that. As a club player and someone who enjoys playing against engines (usually with some type of handicap) I am interested in style and applicability to human games for analysis and playing.
My feeling based open, admittedly few games to date is that stockfish NNUE with the strongest nets seem to be very solid, and perhaps even more human like than classical stockfish? Be interesting to see if style can be fashioned in some manner although this is a very subjective term and perhaps not measurable.
Re: Stockfish NNUE style
Posted: Sun Aug 09, 2020 5:42 pm
by carldaman
That's alright if the preferred style is that of Karpov or Petrosian, but what if one likes Tal better? Besides, too much solidity can be detrimental to a very strong engine, since it can rob it of extra winning chances.
Ideally, the much stronger engine should be inclined to play for a win, rather than not to lose.
Re: Stockfish NNUE style
Posted: Sun Aug 09, 2020 5:56 pm
by Jhoravi
How about her evaluation of French Defense? Will there be improvement considering that NNUE is just trained by regular AB engines?
Re: Stockfish NNUE style
Posted: Sun Aug 09, 2020 6:08 pm
by dkappe
Jhoravi wrote: ↑Sun Aug 09, 2020 5:56 pm
How about her evaluation of French Defense? Will there be improvement considering that NNUE is just trained by regular AB engines?
Except Night Nurse.
Re: Stockfish NNUE style
Posted: Sun Aug 09, 2020 6:54 pm
by Rowen
carldaman wrote: ↑Sun Aug 09, 2020 5:42 pm
That's alright if the preferred style is that of Karpov or Petrosian, but what if one likes Tal better? Besides, too much solidity can be detrimental to a very strong engine, since it can rob it of extra winning chances.
Ideally, the much stronger engine should be inclined to play for a win, rather than not to lose.
I've been lead to believe that nets cant be trained to play in a particular way, to be trained to have a particular style, that in practice they are trained to refute moves, Yet I seem to read information elsewhere that suggests or infers that different styles, can be fashioned?
Re: Stockfish NNUE style
Posted: Sun Aug 09, 2020 7:12 pm
by carldaman
In principle, I can't see why a net can't be trained on the evaluation of a very optimistic and aggressive engine, such as CyberNezh or OpenTal, and then the resulting net would hopefully reflect that style. In practice, things could turn out differently, though.
I hope to eventually learn how to train a net, so I can train one using Nezh, as an experiment.
Nezh is an engine that goes out to play for a win in every game, taking the necessary risks and then some (see link below, as it's only available on lichess).
Re: Stockfish NNUE style
Posted: Sun Aug 09, 2020 7:24 pm
by carldaman
I want to add that it may be more useful, albeit slower, to be able to train against weaker dummy opponents [in separate runs, with White, and then with Black], in order to achieve a desired playing style. This would help circumvent the refutation problem.
Re: Stockfish NNUE style
Posted: Sun Aug 09, 2020 7:28 pm
by ChickenLogic
carldaman wrote: ↑Sun Aug 09, 2020 7:12 pm
In principle, I can't see why a net can't be trained on the evaluation of a very optimistic and aggressive engine, such as CyberNezh or OpenTal, and then the resulting net would hopefully reflect that style. In practice, things could turn out differently, though.
I hope to eventually learn how to train a net, so I can train one using Nezh, as an experiment.
Nezh is an engine that goes out to play for a win in every game, taking the necessary risks and then some (see link below, as it's only available on lichess).
It heavily depends whether you train with a lambda of 1 or less. Lambda 1 means it only tries to predict the evaluation of the position. The closer you set it to zero the more it'll try to learn from the game result (of course some randomness in the data is needed like "temp" for Leela). Getting started with training a net for NNUE is much easier and quicker than training one for Leela. The "Stockfish-Discord" has a lot of good resources on how to train and a dedicated help channel.