TalkChess.com

Just ask on the Leela chess Discord in the #help channel.

https://discord.com/channels/4254194825 ... 1713154049

Lc0 and most of the other NN engines should be run with 2 threads, not 1 (or number-of-GPUs + 1). As far as I know, it's not important, one core is enough. The calculation of the NN engines is done almost exclusively via the GPU. BTW, a very strong Nvidia-RTX2080 was used here... Exactly 2 threads ...

Lc0 and most of the other NN engines should be run with 2 threads, not 1 (or number-of-GPUs + 1).
Incidentally, more than that actually hurts performance (very unlike A/B engines).

A very nice video:
https://www.youtube.com/watch?v=6B0W_LZ57zg

A couple of things. The CCRL Standard Dataset is quite valuable as it provides a trained net benchmark to compare against a net you have trained from the same data. However, as mentioned earlier, the Bad Gyal data will produce stronger nets having more Policy information. And, of course, the actual ...

Don't know what you mean by pre-training, but a suggestion. Doing supervised learning (SL) from games with only the move actually made set to 1 turns out to reduce net strength by about 100-150 Elo. Training games from Lc0 include policy probabilities for all moves and produce stronger nets. It is p...

Sometimes the answer is simply because that's what Alpha Zero did...

Experiment and see what works for you after you have established a working baseline.

Have fun.

Move ordering and pruning is getting to the heart of the individual ecosystem that each engine has, so what works in one may not help at all in another. I've just never found anything other than scanning the current move list subset and picking the highest value to be any improvement. The first move...

Suggest reviewing this first:

https://github.com/Zeta36/chess-alpha-zero

or here for a simpler game like Connect4:

https://github.com/suragnair/alpha-zero-general

the general thinking was that "fancy" sorting does not pay as so much of the tree is pruned in the first place That holds only for the "pick the best of the remaining list and put it to the current place" sort, but not compared to sorting the whole move list up-front with an n^2 algo. All the moves...