I'm doing a small study of the scalability of the reference bot at various numbers of playouts.
I'm also defining level 0 (or zero playouts) as meaning a legal uniformly random move. I still need a lot more games, but in general you eventually start to see a point of diminishing returns for each doubling. I didn't take it any farther than 8192, but my guess is that anything beyond this is not going to give you much. I imagine that it approaches some hypothetical level in an asymptotic way. I may test 16384 later. You also see a big sweet spot between level 32 and 256 where each doubling is worth an enormous improvement. After 256, each doubling gives good but significantly less improvement. Going from 2048 to 4096 gives less than 100 ELO, and after this we get very little with each doubling. Rank Name Elo + - games score oppo. draws 1 refbot-008192 2698 33 32 675 78% 2225 0% 2 refbot-004096 2662 33 32 675 75% 2195 0% 3 refbot-002048 2573 31 31 676 66% 2209 0% 4 refbot-001024 2460 33 33 676 57% 2173 0% 5 refbot-000512 2309 37 38 676 54% 2034 0% 6 refbot-000256 2091 45 47 677 46% 1930 0% 7 refbot-000128 1722 64 65 675 47% 1631 0% 8 refbot-000064 1332 64 62 678 51% 1347 0% 9 refbot-000032 1017 62 63 676 54% 1032 0% 10 refbot-000016 610 48 45 679 55% 780 0% 11 refbot-000008 413 38 37 677 49% 676 0% 12 refbot-000004 255 34 34 676 42% 557 0% 13 refbot-000002 150 33 33 676 34% 530 0% 14 refbot-000001 24 33 34 678 23% 478 0% 15 refbot-000000 0 33 34 676 20% 515 0%
signature.asc
Description: This is a digitally signed message part
_______________________________________________ computer-go mailing list [email protected] http://www.computer-go.org/mailman/listinfo/computer-go/
