=?koi8-r?Q?=F0=C5=D4=D2_=F3=CD=CF=CC=CF=D7?=: 
<[email protected]>:
>I would like to know which program will be stronger: the one with a million 
>pure random games 
>or the one with 10 000 games, but these games will not be pure random...


Uniform random simulations are not enough for the game of Go.  Those 
return just the average over all legal moves of a position but the 
necessary is the maximum value (in nega-max sense) of the position 
with optimal move(s) (though this is practically impossible).  So MCTS 
loves a position with many good moves rather than one with a winning 
and many losing moves.  This is the reason why MCTS is not good to 
find a narrow path to win.

Practically speaking, random simulations have to focus on sampling 
optimal moves.
Hideki
-- 
Hideki Kato <mailto:[email protected]>
_______________________________________________
Computer-go mailing list
[email protected]
http://dvandva.org/cgi-bin/mailman/listinfo/computer-go

Reply via email to