The play-outs are japanese as well while the move selection is probably the same (influenced by the work of Rémi Coulom and the MoGo team) as yours. The urgencies of fields that are not of special interest during the move-selection are biased according to the territory statistics. The urgencies derived from the territory statistics are choosen very close in my case, so that they only have an effect during the progressive widening steps.
I had the same problems you descriped (defending territory) using it to prevent moves on intersections owned by either player. The statistics can only be used near the root. Especially for a better counting after a double pass near the root. The game results obtained this way sometimes differs in some points from the results of GnuGO, but the decision who had won was everytime the same (after more than 500 games played so far). As you surly know, a double pass near the root has a very high influence on the tree-search result, so it's extrem important having correct results there. By the way: a 9x9 CGOS server using japanese rules... I have a dream.. ;) Lars DD> AnchorMan uses that in KGS mode - it will pass quite early sometimes and DD> mark dead stones based on the territory statistics you are talking DD> about. DD> So I assume the play-outs are chinese and the move selection is the same DD> as our bots except you won't move into an intersection that is owned by DD> either player? DD> How reliable is that? I had to be pretty conservative in AnchorMan DD> about using that, it would fail to defend territory unless I made the DD> threshold for ownership pretty high. DD> - Don DD> Lars wrote: >> I had build an Monte-Carlo GO-Engine (GOMonCy) wich uses the Japanese >> scoring system. It reached a win rate against GnuGO 3.6 level 10 of >> stable 50%-52%. I used territorry-statistics about the Monte-Carlo >> outcomes. You get a probability for every field telling you who is the >> owner. It works quite good, but I thougt that nearly everyone is using >> such statistics, isnt't it? Using a threshold to decide that a field >> belongs to a player you can also handle seki situations. Of course, if >> it is losing, the engin will break the seki situation an continue >> losing.. >> >> Am Montag, den 05.11.2007, 16:54 -0800 schrieb [EMAIL PROTECTED]: >> >>> Jason House said: >>> >>>> What about seki situations? >>>> >>>> On Nov 5, 2007 1:41 PM, <[EMAIL PROTECTED]> wrote: >>>> >>>> >>>>> It takes some tricky analysis to work out the Japanese score, due to >>>>> uncertainty about life/death; likewise it's not easy for a program to >>>>> recognize when moving is no longer to its advantage. >>>>> >>>>> How about bringing in a Monte Carlo routine after both players have >>>>> passed?--as a scoring referree, set to fill up the board (but avoiding >>>>> eye-filling >>>>> >>> ------------------------------------------------------------------- >>> -->> and self-atari (except in ko situations) <<-- >>> ------------------------------------------------------------------- >>> >>>>> until all legal >>>>> moves are played... >>>>> >>> ----------------------------------------- >>> This email was sent using AIS WebMail. >>> http://www.americanis.net/ >>> >>> >>> _______________________________________________ >>> computer-go mailing list >>> computer-go@computer-go.org >>> http://www.computer-go.org/mailman/listinfo/computer-go/ >>> >> >> _______________________________________________ >> computer-go mailing list >> computer-go@computer-go.org >> http://www.computer-go.org/mailman/listinfo/computer-go/ >> >> DD> _______________________________________________ DD> computer-go mailing list DD> computer-go@computer-go.org DD> http://www.computer-go.org/mailman/listinfo/computer-go/ _______________________________________________ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
