And COW is up to build 195 as of February 28, 2024. (That's the Macintosh version of Chess Openings Wizard.)
It has Stockfish 16.1 thanks to your reminder about using Homebrew.
www.bookup.com/home/chess-openings-wizard-2016-program-status/
@Toscani
It has Stockfish 16.1 thanks to your reminder about using Homebrew.
www.bookup.com/home/chess-openings-wizard-2016-program-status/
@Toscani
@Toscani said in #20:
> Interesting results...
> www.sp-cc.de/files/statistics_eas_ratinglist.txt
Note that only the engines from "UHO-Top15 Ratinglist" are tested for their "Engine Aggressiveness Score".
Stockfish seems to be the most aggressive according to that list out of the Top 15, but it might not be the most aggressive overall.
@Unseekedspy said in #21:
> Quick question: What is Stockfish 16.1's elo/rating?
That is a very difficult question that has no definitive answer. We tried to explain why in our Frequently Asked Questions:
github.com/official-stockfish/Stockfish/wiki/Stockfish-FAQ#the-elo-rating-of-stockfish
@Timur_chess-2012 said in #25:
> @StockfishNews Has it already been released or is it just being developed?
Stockfish 16.1 was already released. You have the download link at the top of the release notes.
@Freudenfeld said in #30:
> It would be interesting to have more information regarding the two different NNUE files. When or how does Stockfish decide to switch them, IOW when to use the one or the other? Is that fully automated or does it involve custom user configuration?
Stockfish decides automatically when to evaluate with each net. It does so based on a purely materialistic evaluation.
github.com/official-stockfish/Stockfish/blob/e67cc979fd2c0e66dfc2b2f2daa0117458cfc462/src/evaluate.cpp#L180-L201
> As a similar topic, we find in the Lichess options for local SF 16 analysis the option to choose between a 7 MB NNUE or one with 40 MB. How to decide about that?
About Lichess' options, the 7MB NNUE is something special that the official Stockfish 16 itself doesn't use. Its only purpose is to save the user's internet bandwidth when they have to download the engine because 40MB could take a while.
If you want to use the official Stockfish 16, you should choose the 40MB option.
Stockfish 16.1's main network is even bigger plus there is also the smaller network, so Lichess might again decide to create another option for a smaller download size.
As far as I know, the Stockfish 14 available on Lichess also uses a smaller net.
Remember that you can also enjoy faster analysis using Stockfish locally instead of in the browser and also try Stockfish 16.1 before Lichess adds it while still using Lichess itself as an interface using the Lichess Local Engine:
github.com/official-stockfish/Stockfish/wiki/Download-and-usage#free
> Interesting results...
> www.sp-cc.de/files/statistics_eas_ratinglist.txt
Note that only the engines from "UHO-Top15 Ratinglist" are tested for their "Engine Aggressiveness Score".
Stockfish seems to be the most aggressive according to that list out of the Top 15, but it might not be the most aggressive overall.
@Unseekedspy said in #21:
> Quick question: What is Stockfish 16.1's elo/rating?
That is a very difficult question that has no definitive answer. We tried to explain why in our Frequently Asked Questions:
github.com/official-stockfish/Stockfish/wiki/Stockfish-FAQ#the-elo-rating-of-stockfish
@Timur_chess-2012 said in #25:
> @StockfishNews Has it already been released or is it just being developed?
Stockfish 16.1 was already released. You have the download link at the top of the release notes.
@Freudenfeld said in #30:
> It would be interesting to have more information regarding the two different NNUE files. When or how does Stockfish decide to switch them, IOW when to use the one or the other? Is that fully automated or does it involve custom user configuration?
Stockfish decides automatically when to evaluate with each net. It does so based on a purely materialistic evaluation.
github.com/official-stockfish/Stockfish/blob/e67cc979fd2c0e66dfc2b2f2daa0117458cfc462/src/evaluate.cpp#L180-L201
> As a similar topic, we find in the Lichess options for local SF 16 analysis the option to choose between a 7 MB NNUE or one with 40 MB. How to decide about that?
About Lichess' options, the 7MB NNUE is something special that the official Stockfish 16 itself doesn't use. Its only purpose is to save the user's internet bandwidth when they have to download the engine because 40MB could take a while.
If you want to use the official Stockfish 16, you should choose the 40MB option.
Stockfish 16.1's main network is even bigger plus there is also the smaller network, so Lichess might again decide to create another option for a smaller download size.
As far as I know, the Stockfish 14 available on Lichess also uses a smaller net.
Remember that you can also enjoy faster analysis using Stockfish locally instead of in the browser and also try Stockfish 16.1 before Lichess adds it while still using Lichess itself as an interface using the Lichess Local Engine:
github.com/official-stockfish/Stockfish/wiki/Download-and-usage#free
Hello. I dont understand.
ممممممممممممم
Thank you for taking documentation to users seriously. I will run out of excuses not to look into it again.
I am intrigued by the fully NN direction, now added to "using leelas data" for training. I wonder what is left, but the search backbone that was AB and then the forward pruning heuristics.
I wonder where the explanation for the training and leaf evaluations is now explained. If the contingencies of using any type of leaf evaluations is explain as before.
I remember reading in some issue that there was still some "simple evaluation", and that there was material count involved.
So is that to say that the hand-crafted evaluation has been brought back to its very original algebraic form, and the more empirical with hand-crafted input features NN bases functions are there to let some supervised or reinforcement learning do the appropriate weight of all those input features per the training/testing/self-play (or) environment feedback (however constructed).. sorry for not trying to explain. but if someone is ML trained, they might decrypt my text, I hope.
Edit: looking at some outlinks. The architecture. Nice attempt. Font problems. Looking for outlinks to the ML information flow documentation (that is not about the architecture novelties but about specifying the training set up. In SL it means the input data "vector", and the target data vector over some training set of positions, and given target scores from whatever source, used to be SF itself in the past. In RL, I know how A0 and LC0 might do it, but how does SF do it. I have read about using MCTS in some SF related engines (same type A descendant), fire??? something, and komodo/dragon. That type of documentation would be readable by users seeking to interpret what SF might be doing to their chess. It would be actually less technical hurdle, as such methods are inspired from how we learn. The box and arrow, with no font problems, would be easier to read. and create. I actually secretly hope that what I describe is actually out there, but I have been burned in the past, and did not keep an open mind to keep looking for it, and now I am missing it.... pointer please, then. anyone. to that specifically. I got lazy.
I am intrigued by the fully NN direction, now added to "using leelas data" for training. I wonder what is left, but the search backbone that was AB and then the forward pruning heuristics.
I wonder where the explanation for the training and leaf evaluations is now explained. If the contingencies of using any type of leaf evaluations is explain as before.
I remember reading in some issue that there was still some "simple evaluation", and that there was material count involved.
So is that to say that the hand-crafted evaluation has been brought back to its very original algebraic form, and the more empirical with hand-crafted input features NN bases functions are there to let some supervised or reinforcement learning do the appropriate weight of all those input features per the training/testing/self-play (or) environment feedback (however constructed).. sorry for not trying to explain. but if someone is ML trained, they might decrypt my text, I hope.
Edit: looking at some outlinks. The architecture. Nice attempt. Font problems. Looking for outlinks to the ML information flow documentation (that is not about the architecture novelties but about specifying the training set up. In SL it means the input data "vector", and the target data vector over some training set of positions, and given target scores from whatever source, used to be SF itself in the past. In RL, I know how A0 and LC0 might do it, but how does SF do it. I have read about using MCTS in some SF related engines (same type A descendant), fire??? something, and komodo/dragon. That type of documentation would be readable by users seeking to interpret what SF might be doing to their chess. It would be actually less technical hurdle, as such methods are inspired from how we learn. The box and arrow, with no font problems, would be easier to read. and create. I actually secretly hope that what I describe is actually out there, but I have been burned in the past, and did not keep an open mind to keep looking for it, and now I am missing it.... pointer please, then. anyone. to that specifically. I got lazy.
It's lovely
“The free and open source engine that you love, now greater and more terrible than ever before”, huh :-)
Stockfish 16.1 has 27 Elo points more than stockfish 16. But on the ccrl 40/15 and ccrl blitz site (therefore playing with other engines), it's weaker than stockfish 16