LMR testing

[gaard]

Testing LMR chhanges at 5"+0.05" is a very different beast versus testing at 60"+0.6". What is the longest time control you find useful to validate said changes? I usually have to increment time by 30x but there must be something more efficient, or some alternate method?

To add some context, I'm referring to the usual LMR formula. log2(min(moveNum, 64)) * log2(min(depth, 64)) * X1 + X2

[Steve Maughan]

Hi Martin,

Interesting — I had thought basically any time control would work. I'll do some digging.

As for the "usual LMR formula", do you know which engines us this formula? Does Stockfish use it?

— Steve

[Ciekce]

As for the "usual LMR formula", do you know which engines us this formula? Does Stockfish use it?

stockfish has a variation on it, it's almost a standard now - you can see it in SP (with unclamped moves and depth) here https://github.com/Ciekce/Stormphrax/bl ... le.cpp#L37

pick any strong engine off CCRL and chances are it'll use that formula or something very similar

[gaard]

Hi Martin,

Interesting — I had thought basically any time control would work. I'll do some digging.

As for the "usual LMR formula", do you know which engines us this formula? Does Stockfish use it?

— Steve

Berzerk: log(d) * log(m) * x + y
Demolito: log(d) * x + log(m) * y
Senpai: log2(d) * log2(m) * x
Ethereal: log(d) * log(m) * x + y

There is some variation, but more-or-less, the formulas are near identical

[Steve Maughan]

That's great! Thanks! I have come across this formula but didn't appreciate how widespread it was being used.

— Steve

[gaard]

Hi Martin,

Interesting — I had thought basically any time control would work. I'll do some digging.

As for the "usual LMR formula", do you know which engines us this formula? Does Stockfish use it?

— Steve

Changes that can significantly affect the tree or branching factor seem to be the most sensitive to time control in my experience. They could yield more Elo at shorter tcs, but turn negative at longer, and vice-versa. Usually my slowest tc testing is 60"+0.6", only 6x slower than my fastest. Maybe I need a greater breadth?

[Bart Weststrate]

Testing LMR chhanges at 5"+0.05" is a very different beast versus testing at 60"+0.6". What is the longest time control you find useful to validate said changes? I usually have to increment time by 30x but there must be something more efficient, or some alternate method?

To add some context, I'm referring to the usual LMR formula. log2(min(moveNum, 64)) * log2(min(depth, 64)) * X1 + X2

What is X1 and X2? Maybe a stupid question...

[Ciekce]

Testing LMR chhanges at 5"+0.05" is a very different beast versus testing at 60"+0.6". What is the longest time control you find useful to validate said changes? I usually have to increment time by 30x but there must be something more efficient, or some alternate method?

To add some context, I'm referring to the usual LMR formula. log2(min(moveNum, 64)) * log2(min(depth, 64)) * X1 + X2

What is X1 and X2? Maybe a stupid question...

Tunable params. X1 = 1/2.36 and X2 = 0.77 in my engine

[Bart Weststrate]

Thanks.

[Frank Quisinsky]

BTW:
Wow ... Bart Westrate is here!
Jan Louwman gave often the information: Clearly the most talented chess programmer (middle of the 90er).
Kallisto 1.83 was most aggressive ... produced the most short wins!
OK, newer versions are stronger and the DOS graphic from Kallisto was a mega-event.

Bart, today I have excactly the same fun with your Kallisto II as in the time you gave me the program in Leiden.

Send shortly my regards!

Best
Frank

Maybe you have interest to create an UCI engines for us!
Sure, an event!