Werewolf wrote:Would you be able to estimate the elo gain per doubling the number of rollouts?
Is it anywhere near a normal engine where doubling the amount of time yields around 50 elo?
It is probably more, but so far no-one has calculated a scaling curve for Leela Chess. In theory, NN based engines scale better than Alpha-Beta engines.It would be a very good idea to test that though, would you be interested?
I'm interested in helping the project, yes. I was planning on donating GPU time. But I'm also running test matches against a dedicated unit as a benchmark of how strong LCZ is.
At this point, enter "go" to prompt it to play a move. If that works, the rest is a matter of configuring the command line options in Arena properly, I saw some answers to that in this thread (or the predecessor thread anyway)
I also followed all the instructions, but I'm getting an error: "lczero.exe has stopped working"
If you have cutechess_cli or some other software that can conduct chess engine tournaments, could you run a competition between the CPU and GPU versions at equal playouts? There has been talk on Github that GPU version is producing large error due to OpenCL calculations sometime, it would be cool if you could provide data on that (and share on Github).
Your tournament so far is a good start, but just means that this GPU just doesn't work well for you. We'd need equal playout settings (e.g. 800) to make this meaningful in evaluating whether the GPU calculation errors hurt playing strength a lot.