New net: The White Rose

[Damir]

Vielen dank für die Weisse Rose Netz Dietrich.. Kannst Du uns ein bisschen mehr erzählen über dieses Netzwerk ?

[dkappe]

Vielen dank für die Weisse Rose Netz Dietrich.. Kannst Du uns ein bisschen mehr erzählen über dieses Netzwerk ?

Also, die Daten für das Netz sind ähnlich wie Night Nurse — mcts/nn. Night Nurse kriegt aber nur 800 nodes pro Zug während Dark Horse 1600 und noch mehr bekommt. Bei „The White Rose“ sind noch ein Paar weitere Klimmzüge beim Training dabei.

[purechess]

It cannot be embedded in and distributed with a GPL work

Finally you added a license and you have chosen https://creativecommons.org/licenses/by-nd/4.0/deed.en

That the net can't be embedded in GPL software is clear.
But it can be loaded by the engine as binary I guess when distributed ?!

[dkappe]

It cannot be embedded in and distributed with a GPL work

Finally you added a license and you have chosen https://creativecommons.org/licenses/by-nd/4.0/deed.en

That the net can't be embedded in GPL software is clear.
But it can be loaded by the engine as binary I guess when distributed ?!

You can distribute it alongside a nnue capable engine with the license text, just not embed it in a GPLv3 engine. That would violate the terms of the GPL.

[carldaman]

I’ve approached the training with Dark Horse data a little differently. It’s a little under inflated, but not as much as DH-0.2.

https://www.patreon.com/posts/47771982

Note that this net is distributed under the Creative Commons attribution, no derivatives license. As such it cannot be embedded and distributed in a GPLv3 work or others with an incompatible license.

Thanks for your efforts, but The Baron 3.441, clearly rated below 3000 CCRL, can seemingly draw at will with White against recent Cfish + White Rose net, at rapid TCs + no book. A massive fail, as far as I'm concerned, but this deficiency is also shared with some of the latest SF NNUE nets.

I frequently like to test against much weaker engines and without a book, things that relatively very few others do. This can bring to light such aberrations. Of course, these overly drawish tendencies from the NNUE Black side is something that became apparent soon after NNUE originally came out. The absence of a contempt setting combined with risk-averse play from the Black side continues to be a general problem with NNUE.

I will test some of your other recent nets against The Baron, under the same conditions, and report back my impressions.

Interesting. Can we expect The Baron in the SuFi soon?

Not any time soon, I'm afraid, but I guess that's point.

I'll test the DH2 net next today, the most aggressive of your nets in my experience. Out of curiosity, I also discovered that FF2+freenet and ShashChessNNUE+Tal personality can crush the Baron with very enterprising play from the Black side at 10 m + 8 s rapid.

[dkappe]

Not any time soon, I'm afraid, but I guess that's point.

I'll test the DH2 net next today, the most aggressive of your nets in my experience. Out of curiosity, I also discovered that FF2+freenet and ShashChessNNUE+Tal personality can crush the Baron with very enterprising play from the Black side at 10 m + 8 s rapid.

I’ve been playing around with lambda in the training, which is the proportion of game outcome vs eval. Lambda of 1.0 is all eval. I’ve found that a lambda of 0.5, so 50-50 eval and result, yields nets that are better against the top sf nets but at the cost of conservative play and an ungodly number of draws. More eval (I think DH-0.2 was 0.9?) gives you a more entertaining and aggressive style. So it’s a trade off.

[carldaman]

Not any time soon, I'm afraid, but I guess that's point.

I'll test the DH2 net next today, the most aggressive of your nets in my experience. Out of curiosity, I also discovered that FF2+freenet and ShashChessNNUE+Tal personality can crush the Baron with very enterprising play from the Black side at 10 m + 8 s rapid.

I’ve been playing around with lambda in the training, which is the proportion of game outcome vs eval. Lambda of 1.0 is all eval. I’ve found that a lambda of 0.5, so 50-50 eval and result, yields nets that are better against the top sf nets but at the cost of conservative play and an ungodly number of draws. More eval (I think DH-0.2 was 0.9?) gives you a more entertaining and aggressive style. So it’s a trade off.

The good news is that both DH2 and NiNu4 were able to win with Black. However, only DH2 managed to totally outplay the much weaker Baron in the middlegame. NiNu4 got a slight structural advantage into the endgame, and was only able to grind out a win after move 40. This confirms my earlier impressions about the DarkHorse nets. Sample size is minimal, but I can tell a lot just by watching a few games, especially when an overarching pattern emerges in the style of play.

Playing strength (for these nets) is so enormous these days, that I'm always happy to trade some strength for style. It's good that you're training different nets and you're not committed one way or another - this way you can pursue both strength and style in parallel. Thumbs up on that!

[dkappe]

Not any time soon, I'm afraid, but I guess that's point.

I'll test the DH2 net next today, the most aggressive of your nets in my experience. Out of curiosity, I also discovered that FF2+freenet and ShashChessNNUE+Tal personality can crush the Baron with very enterprising play from the Black side at 10 m + 8 s rapid.

I’ve been playing around with lambda in the training, which is the proportion of game outcome vs eval. Lambda of 1.0 is all eval. I’ve found that a lambda of 0.5, so 50-50 eval and result, yields nets that are better against the top sf nets but at the cost of conservative play and an ungodly number of draws. More eval (I think DH-0.2 was 0.9?) gives you a more entertaining and aggressive style. So it’s a trade off.

The good news is that both DH2 and NiNu4 were able to win with Black. However, only DH2 managed to totally outplay the much weaker Baron in the middlegame. NiNu4 got a slight structural advantage into the endgame, and was only able to grind out a win after move 40. This confirms my earlier impressions about the DarkHorse nets. Sample size is minimal, but I can tell a lot just by watching a few games, especially when an overarching pattern emerges in the style of play.

Playing strength (for these nets) is so enormous these days, that I'm always happy to trade some strength for style. It's good that you're training different nets and you're not committed one way or another - this way you can pursue both strength and style in parallel. Thumbs up on that!

I’m training Dark Horse 0.3 right now. Same data as The White Rose, but 100% eval. Hopefully it gives you the style you so enjoy.

[carldaman]

Fantastisch! Got my fingers crossed - hope it's a good one.

[dkappe]

Fantastisch! Got my fingers crossed - hope it's a good one.

Well, definitely not as strong as TWR, but hopefully more aggressive. Dark Horse 0.3. https://www.patreon.com/posts/dark-horse-0-3-47913036