Here is a thread in which I gave some examples (to 1 and 2 threads, I think). I think I had some broader rules o thumb written here somewhere later, but I couldn't find them quickly.jp wrote: ↑Sat Jul 20, 2019 8:13 pmCan you give a rough rule of thumb for how much non-determinism there is (for SF, etc.) as a function of number of threads and number of nodes per move?Laskos wrote: ↑Sat Jul 20, 2019 7:25 pm Well, in Lc0 they do introduce a non-zero temperature for some number of moves (maybe 20?) during the training, some sort of noise obeying IIRC the Dirichlet distribution with some K. We can randomize similarly Lc0 via the UCI settings. I do this when I have few or single starting position, for say 4 moves. It can actually produce a good opening tree with well chosen parameters. But I haven't even checked if Lc0 with no noise on one of thread isn't deterministic to fixed nodes. With default settings (0 temperature and 2 threads), it appers to be no less deterministic than SF or Komodo on two threads.
I am not sure why in AB engines it's hard to introduce a simple noise like Rybka did yers ago or Komodo nowadays. It seems no to be computationally intensive.
In the other thread, Uri believes there is a lot of variety, while I worry it has almost none (for purposes where we want a lot).
e.g. viewtopic.php?p=805353#p805353
viewtopic.php?f=2&t=61462
I tend to agree with Uri that a multi-threaded SF is somewhat more random than Leela on 2 threads (temperature=0).