JohnS wrote: ↑
Wed Oct 09, 2019 6:16 am
MikeB wrote: ↑
Tue Oct 08, 2019 7:42 pm
well if you have a macOS computer anyway , or if you can compile your own
other OS flavors will be released upon receipt.
all versions have all bell and whistles, including play by FIDE Elo (or CCRL Elo) and improved adaptive and variety play
FD versions have 'fortress detection' code written Joe Ellis.
Bluefish and Blue-Honey versions have Bluefish evaluation code written by Eelco de Groot.
that's right - 8 stockfish derivative flavors , including
there's build script in the src folder, on my machine it takes 62 seconds to build all 8 flavors, just over two minutes for PGO flavors ( all macOS versions were PGO optimized)
why "X" ? , it signifies that this version is related to Stockfish 10 and development code added since the release of Stockfish 10.
Excellent work Mike!
Can you please clarify the use of adaptive, play by Elo and variety? Do adaptive and play by ELo override the strength setting? Can they be used together (does this make any sense) or is it one or the other?
What does variety do and how does it interact with adaptive and play by ELo?
Can you provide examples of typical uci settings for lower strength options?
By the way, Leela scoring is now the default and it initially confused me because a score of +50 in the initial position looked crazy (you finally proved chess is a forced win for White
). I personally think centipawns should be the default as that is what most of us are used to.
Many thanks for a great engine.
Excelled questions. - first play by Elo is nps based - you should use the the uci sleep option when using any weakening level as the program will move instantly without it checked. There is no randomization except at the weakest levels. Why? Well Stockfish is so strong that even at 50 nps it plays in the 1500 to 1600 range (FIDE Elo with CCRL that’s about 1200) - the only way to get it below that Elo is to introduce randomness.
Variety play introduces randomness across the entire Elo spectrum. It costs about 300 Elo. To make up for that ,when it is used with play by Elo , I increase the nps, so it will still play close the Elo selected.
Adaptive play also introduces variety but does it differently. The code used here is from “EasyPlay”. For most people it will have range of playing of about 300 Elo as it will purposely throw you a bone when it gets ahead. Like variety play , the nps is increased under the hood , to keep the Elo close to what you selected.
You can tell when an adaptive play move is used , as the move played will not be in sync with the pv. So when you see that , that’s your chance to get back in the game. ;>)
With variety play , the pv will be in sync with the pv. They both can be used together and the play will be slight weaker than using one or the other - although the nps is raised even higher.
At higher levels of play , the adaptive play has even a wider range of playing - if you play at Elo 2900 with adaptive play , it will play close to Crafty. With a weaker opponent it could play as low 2300 to 2400 (CCRL rating), Variety play does not have a plying range capability - it plays at one level, just weaker than before.
On the LC0 scoring % , that is set to my preference, it’s what I use for all my analysis for my games and I am my number one customer ;> ) - for me that is more useful and intuitive than centipawn as it provides you the expected outcome of the game which centipawn does not provide. Keep in mind , the internal scoring mechanism is still the same - it is just converted using a formula. But I recognize that I am in the minority so there is a toggle to switch it back to centipawn scoring and with most good GUIs , one can make that change persistent. One thing I like about it is that it always real clear that are you looking at the game from my the side to move viewpoint so that box should be checked on your GUI - not from Black or white side viewpoint.
It’s one of those things, the more you use it , the more you appreciate it. Now if somebody has maintained a record of thousands and thousands of positions that they scored from a centipawn vantage point - well obviously it will not be for them.
But for actually playing games or doing analysis for a correspondence game - I like it better. I focused a lot harder on games where I am under 50% ;>).