You are faster than me
The windows compile should be there now. (I had to update a few things before compiling for Windows, which takes ages...)
Have fun with it
Moderators: hgm, Rebel, chrisw
You are faster than me
Thanks, Richard.abulmo2 wrote: ↑Thu Dec 16, 2021 10:16 pmYou are faster than me
The windows compile should be there now. (I had to update a few things before compiling for Windows, which takes ages...)
Have fun with it
Code: Select all
Score of Princhess-0.6.0 vs Princhess-0.5.1: 524 - 356 - 120 [0.584] 1000
... Princhess-0.6.0 playing White: 273 - 172 - 55 [0.601] 500
... Princhess-0.6.0 playing Black: 251 - 184 - 65 [0.567] 500
... White vs Black: 457 - 423 - 120 [0.517] 1000
Elo difference: 58.9 +/- 20.4, LOS: 100.0 %, DrawRatio: 12.0 %
"Switch to neural net for eval"Lana Samson wrote: ↑Fri Dec 17, 2021 8:28 am Princhess 0.6.0 released
https://github.com/princesslana/prinche ... /tag/0.6.0
This is mainly based on timing (release what I have before I take off for Christmas). For me, she seems to play more inconsistently than previously, but has gained elo in self play. Not sure how she'll do vs other engines.
I disagree. As long as it is the REAL Hiarcs (not like "Fritz" which has had 20 authors since the legendary Frans Morsch) and retains its nice positional playing style, I will buy it in an instant.
Yeah, granted. Deep Fritz 13 was the last real Fritz. Nowadays "Fritz" is just the user interface program, running some engine renamed "Fritz". HIARCS 14.1 already was 3150 Elo or thereabouts, on 4 threads. It's possible to reach 3100 on a single thread (see Strelka), gain another 100-150 for multithreading on 4 cores, which makes 3200-3250. Then you get into the "realm of the weird", where you need to do lots of work for every 20-30 Elo. (Stockfish has been at it for what; 10 years?) If Mark Uniacke did this on his own, I can imagine that it takes 10 years. I can also imagine that he at some point decided to "just" outfit HIARCS with a NNUE network of his own making (keeping the 14.1 evaluation as an option) and then release it as HIARCS 15. It would make sense.BrendanJNorman wrote: ↑Fri Dec 17, 2021 12:27 pm I disagree. As long as it is the REAL Hiarcs (not like "Fritz" which has had 20 authors since the legendary Frans Morsch) and retains its nice positional playing style, I will buy it in an instant.
I'm sure a lot of old-school computer chess fellas fall into the same boat as me.
Nostalgia is a hell of a drug.
Previously Princhess used a ML learned linear model, now uses a ML learned net. In both cases the model is included in the exe.Gabor Szots wrote: ↑Fri Dec 17, 2021 9:47 am"Switch to neural net for eval"Lana Samson wrote: ↑Fri Dec 17, 2021 8:28 am Princhess 0.6.0 released
https://github.com/princesslana/prinche ... /tag/0.6.0
This is mainly based on timing (release what I have before I take off for Christmas). For me, she seems to play more inconsistently than previously, but has gained elo in self play. Not sure how she'll do vs other engines.
Will you please elaborate on that? I can see the exe size has not changed so it must be some home-cooked net.
This is expected. The source just has a dummy/dev version, with releases setting the version when they're built.BTW, source compiles into 0.0.0-dev, not 0.6.0.
How Strong is Berserk 7.0 NNUE https://chessengines.blogspot.com/2021/ ... -nnue.htmlLana Samson wrote: ↑Fri Dec 17, 2021 4:56 pmPreviously Princhess used a ML learned linear model, now uses a ML learned net. In both cases the model is included in the exe.Gabor Szots wrote: ↑Fri Dec 17, 2021 9:47 am"Switch to neural net for eval"Lana Samson wrote: ↑Fri Dec 17, 2021 8:28 am Princhess 0.6.0 released
https://github.com/princesslana/prinche ... /tag/0.6.0
This is mainly based on timing (release what I have before I take off for Christmas). For me, she seems to play more inconsistently than previously, but has gained elo in self play. Not sure how she'll do vs other engines.
Will you please elaborate on that? I can see the exe size has not changed so it must be some home-cooked net.
This is expected. The source just has a dummy/dev version, with releases setting the version when they're built.BTW, source compiles into 0.0.0-dev, not 0.6.0.
Do you mean using Berserk's net in Princhess? Princhess doesn't support using other nets (or even any net apart from the one compiled in yet)Chessqueen wrote: ↑Fri Dec 17, 2021 5:26 pmHow Strong is Berserk 7.0 NNUE https://chessengines.blogspot.com/2021/ ... -nnue.htmlLana Samson wrote: ↑Fri Dec 17, 2021 4:56 pmPreviously Princhess used a ML learned linear model, now uses a ML learned net. In both cases the model is included in the exe.Gabor Szots wrote: ↑Fri Dec 17, 2021 9:47 am"Switch to neural net for eval"Lana Samson wrote: ↑Fri Dec 17, 2021 8:28 am Princhess 0.6.0 released
https://github.com/princesslana/prinche ... /tag/0.6.0
This is mainly based on timing (release what I have before I take off for Christmas). For me, she seems to play more inconsistently than previously, but has gained elo in self play. Not sure how she'll do vs other engines.
Will you please elaborate on that? I can see the exe size has not changed so it must be some home-cooked net.
This is expected. The source just has a dummy/dev version, with releases setting the version when they're built.BTW, source compiles into 0.0.0-dev, not 0.6.0.