search

lightvector / KataGo

GTP engine and self-play learning in Go
https://katagotraining.org/
Other
3.57k stars 564 forks source link

Can 20-blocks-100-playouts win a 8D human? #465

Open oycq opened 3 years ago

oycq commented 3 years ago

Can 20-blocks-100-playouts win a 8D human?

Friday9i commented 3 years ago

Hard to be sure, but latest 20b are very strong, so I would definitely answer yes. And to the opposite question "Can a 8D human win latest 20-blocks-100-playouts?", which is a very different question, I'm not sure of the answer: I would say "hardly"... That should be tested to get some confidence in these answers. Anyone has some experience on that?

oycq commented 3 years ago

thank you

oycq commented 3 years ago

Terrific answer!

oycq commented 3 years ago

And another question, how about 10-playouts? Can we say that (100-playouts / 10-playouts) == (8D / 7D)？ or （8D / 6D）

Friday9i commented 3 years ago

IIRC, KataGo 40b with 1 visit was about 7D or 8D on KGS... And 20b with 10 visits should be about equivalent to 40b 1 visit (there is ~400 Elo difference between best 20b and best 40b at visit parity. And 10x difference in visits is about 500 or 600 Elo difference, but this rule works for > 10 visits, it's less below 10 visits, so possible 300 to 500 Elo ==> 20b 10 visits is probably about equivalent to 40b 1 visit) So I would say 20b 10 visits should be roughly about 7D or 8D KGS (give this estimate at least 1 rank of uncertainty, that's an order of magnitude)