peter wrote: ↑Sat Oct 03, 2020 1:32 pm
Thanks, Andrea!
What does MCTS checked do?
Looking forward to testing regards
The MCTS is based on the original Sthepane Nicolet idea (https://github.com/snicolet/Stockfish/t ... ecarlo/src).
Clearly, we can't on a cpu, use the MCTS on all threads, but, as for the net, we think we can effectively use a "lighter" version.
In this case, only one thread (the first not main) uses the MCTS.
On a single thread mode, there is no MCTS.
In this manner, the first tests show a very interesting and promising behavior.
Every test, ideas and help is truly appreciated: this is the goal of open source projects
Sorry, and perhaps it is just me (totally low-tech concerning engine performance) but I always have trouble understanding your initial 'explanations'.
Wording like this confuses me: "In this case, only one thread (the first not main) uses the MCTS.
On a single thread mode, there is no MCTS."
So, let me just ask....from Komodo MCTS, I thought the ideas was to combine PV/Threads in a way that - for example, lets say look at 4 PV using 4 threads so that each PV gets to work on its own and create better analysis. Is that how ShashChess 15 works? If so, for a given position, can each 'personality' (algorithm) express itself on each of those 4 PV?
And do this while using NNUE?
If so, this seems truly remarkable! Indeed, rather what I have been expecting Komodo 15 to look like.
Cornfed wrote: ↑Sat Oct 03, 2020 6:47 pm
Sorry, and perhaps it is just me (totally low-tech concerning engine performance) but I always have trouble understanding your initial 'explanations'.
Wording like this confuses me: "In this case, only one thread (the first not main) uses the MCTS.
On a single thread mode, there is no MCTS."
So, let me just ask....from Komodo MCTS, I thought the ideas was to combine PV/Threads in a way that - for example, lets say look at 4 PV using 4 threads so that each PV gets to work on its own and create better analysis. Is that how ShashChess 15 works? If so, for a given position, can each 'personality' (algorithm) express itself on each of those 4 PV?
And do this while using NNUE?
If so, this seems truly remarkable! Indeed, rather what I have been expecting Komodo 15 to look like.
This simply means, on a thread, the search is conducted by the Montecarlo algorithm.
So, in this case, there is (at least at the moment) no Shashin theory.
The other threads use classical Alpha/Beta and for them there is Shashin theory.
All search is conducted like this: PV or not.
Cornfed wrote: ↑Sat Oct 03, 2020 6:47 pm
Sorry, and perhaps it is just me (totally low-tech concerning engine performance) but I always have trouble understanding your initial 'explanations'.
Wording like this confuses me: "In this case, only one thread (the first not main) uses the MCTS.
On a single thread mode, there is no MCTS."
So, let me just ask....from Komodo MCTS, I thought the ideas was to combine PV/Threads in a way that - for example, lets say look at 4 PV using 4 threads so that each PV gets to work on its own and create better analysis. Is that how ShashChess 15 works? If so, for a given position, can each 'personality' (algorithm) express itself on each of those 4 PV?
And do this while using NNUE?
If so, this seems truly remarkable! Indeed, rather what I have been expecting Komodo 15 to look like.
This simply means, on a thread, the search is conducted by the Montecarlo algorithm.
So, in this case, there is (at least at the moment) no Shashin theory.
The other threads use classical Alpha/Beta and for them there is Shashin theory.
All search is conducted like this: PV or not.
To give a real world example:
I have ShashChess 15 loaded with the following in Chessbase .
Parameters (properties):
3 threads
MCTS checked
EvalFile: nn-baeb9ef2d183.nnue
and I have the Chessbase gui set to look at 3 pv
Position at move 8 after a capture on e2; 41 ply:
1 -- 8.Qxe2 -.90
2 -- 8.Nxe2 -1.17
3 -- 8.Bxe2 - 1.47
Should I be using 4 threads instead of 3 (when looking at 3 pv)? Otherwise, should I be using ShashChess in some other way? Obviously I am only interested in ShashChess for analysis, not 'game play'.
Thanks for your time explaining!
amchess wrote: ↑Sat Oct 03, 2020 9:00 pm
Yes, even at pv 3, you can continue to use 4 threads
And so the 4th thread would be used for the NNUE...which actually influences the 3 PV and their evaluations viewed in the GUI. I think that is what you are saying and if so...pretty nice addition to ShashChess!
peter wrote: ↑Sat Oct 03, 2020 1:32 pm
Thanks, Andrea!
What does MCTS checked do?
Looking forward to testing regards
The MCTS is based on the original Sthepane Nicolet idea (https://github.com/snicolet/Stockfish/t ... ecarlo/src).
Clearly, we can't on a cpu, use the MCTS on all threads, but, as for the net, we think we can effectively use a "lighter" version.
In this case, only one thread (the first not main) uses the MCTS.
On a single thread mode, there is no MCTS.
In this manner, the first tests show a very interesting and promising behavior.
Every test, ideas and help is truly appreciated: this is the goal of open source projects
So it is more efficent in 2 thread tournaments. Can you put an option for "pure mcts" uses only mcts in every thread?