That's all fine, but UCT with exploration term == 0 does not asymptotically converge in a ternary-valued space.
The problem is that the search is closing its eyes to the possibility of winning. Both sides are happy to do that because neither side is exploring. -----Original Message----- From: [email protected] [mailto:[email protected]] On Behalf Of Willemien Sent: Friday, April 01, 2011 9:30 AM To: [email protected] Subject: Re: [Computer-go] 7.0 Komi and weird deep search result On Fri, Apr 1, 2011 at 9:28 AM, <[email protected]> wrote: > Quoting Michael Williams <[email protected]>: > >> I wonder what would happen if you counted jigo as a loss for both >> players... > > Interesting, it would then avoid all lines that are likely to lead to a > draw. The problem is that it then not be able to see any difference between > a jigo and a loss which may be a problem in a position which is a clear > jigo. What happens deeper in the tree does not matter, if there is only a > few jigo candidate moves and many losing moves, the program will often pick > a losing move. > > So I think jigo must be evaluated at least better than losing perhaps > something like 25% or so. But then I need to program some more because it > would change how I update the tree, and that part of the program is quite > messy with all the AMAF updating and so on. > > The question is: should jigo be counted balanced for both players or should > one perhaps bias the randomized winner for jigo playout. For example if the > program is winning 60% at root then jigo would count as win in only 40% of > the cases for the color of program in the playouts. This would perhaps drive > the program more towards winning lines when it wins and towards jigo when it > loses. > If you follow my idea : - count a win as 2 W - count a loss as 2 L this gives the option to - count a jigo as 1 W and 1 L the winrate with jigo will go to 50%. _______________________________________________ Computer-go mailing list [email protected] http://dvandva.org/cgi-bin/mailman/listinfo/computer-go _______________________________________________ Computer-go mailing list [email protected] http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
