KataGo Parameters

[RobertJasiek]

Cache and Mutex

With 64 GB of system RAM, what values should I set instead of these defaults?

nnCacheSizePowerOfTwo = 20
nnMutexPoolSizePowerOfTwo = 16

I think that nnCacheSizePowerOfTwo = 23 results in up to ~25 GB cache so would better use my 64 GB RAM, or not? Accordingly, I guess that nnMutexPoolSizePowerOfTwo = 19 would fit to the other adjusted value.

However, if I let KataGo analyse one position for hours and want to use and fill my 64 GB RAM (except for system usage etc.), will KataGo fill it automatically anyway regardless of these two parameters or what do I need to set to enable full RAM filling, when needed?

What is mutex?


Temperature

If I want more randomisation of the first 4 moves of the game, what values should I set for the following parameters?

# chosenMoveTemperatureEarly = 0.5
# chosenMoveTemperatureHalflife = 19
# chosenMoveTemperature = 0.10


LCB

What is LCB in useLcbForSelection?

[Yakago]

I didn't find the discussions in Leela/KataGo about when LCB was implemented, but here is a reference to what it is:
http://www.cs.cornell.edu/courses/cs478 ... otes15.pdf
https://www.geeksforgeeks.org/upper-con ... -learning/

[RobertJasiek]

Thanks for the links! I guess we should treat useLcbForSelection like one of the internal parameters and better not mess with it?

[Yakago]

From what I understand/recall, it has been found to generally improve move selection (higher rating in internal tests with the setting enabled)

I wouldn't think of it as something to optimize and would leave it on unless someone gives a good reason as to why not.


Edit: https://github.com/leela-zero/leela-zero/issues/2411 - here is discussion from 2019 about it and whether or not its good to enable in training games. A quote from the first post (by lightvector) is
"So it seems like at least in KataGo, LCB seems not to be a clear benefit for training, and if anything, LCBMoves perhaps makes it worse. This is despite the fact that I've already verified with the same versions of KataGo that LCB easily adds +100 Elo or more at similar numbers of visits!"

Where we see that the parameter appears to be very beneficial in terms of playing strength, but may not be ideal for producing training games