Álvaro Begué: <[email protected]>: >Don, I assume you mean "with high handicap games...". The problem is >that dynamic komi assumes there will be points to be gained later in >the game, and the program might be happy to get into a situation where >the opponent has a lot of safe territory, and then there are no >opportunities to recover from that.
I think the assumption of dynamic komi is "playing my best for even games is also the best for handcap games." The problem of Zen in high handicap games could be the estimated score is not accurate enough and cannot be used wholly (ie, multiplying by a factor smaller than 1) to avoid unstable play. In fact, I've been working on improving the play of Zen in high handicap games recently and found the scores of simulations have much larger variations than expected. I'm now seeking better way to get more accurate average than flat (simple) avaraging of all scores of simulations. Priortizing the scores of the simulations started from deeper nodes in the tree, for example. # Using some feedback mechanism to keep WR around 50% (used in FudoGo and pachi?) is an another issue. Hideki >We could do something different for handicap games: When it's the >weaker player's turn, instead of using the heavy playout policy all >the time, revert to the light playout policy (random except for not >playing in eyeish points) with some probability p. This way the >program will be happy to leave areas of the board undecided, with the >hope that the weaker player won't know exactly how to defend them. The >probability p of picking from the light playout policy can be adjusted >dynamically as well. I am pretty sure I didn't come up with this, so >perhaps someone has tried it and can tell me why it doesn't work. :) > >Álvaro. > > >On Fri, Jun 17, 2011 at 4:37 PM, Don Dailey <[email protected]> wrote: >> With high komi games you are starting the game from a dead lost position so >> you almost HAVE to assume your opponent is stupid and take some "unsound" >> risk. Of course risk is not "unsound" if you are losing anyway. I see >> no problem with your idea but the devil is in the details. >> >> On Fri, Jun 17, 2011 at 4:23 PM, Stefan Kaitschick >> <[email protected]> wrote: >>> >>> Zen19S is an account on KGS with long time controls(20 + 30/5)*, running >>> on acluster of 6 pcs. It holds a solid 4dan rating. >>> I think it's handicap openings have really improved with both black and >>> white, and I think dyn. komi is a big part of this. >>> But I have seen some 6 stone games as white(the highest number for rated >>> games), that are quite insipid. >>> Zen just lets black take 2 60 point corridors on each side, and cannot >>> compensate in the center. >>> I think this happens, because early in the game, when dyn. komi is high, >>> Zen forsakes the option of future side invasions, >>> and later, when dyn. komi is sinking, it has no recourse, because the game >>> has already been decided. >>> So here's my bright idea: how about modeling opponent incompetence >>> directly? >>> This would only work as white ofcourse. It would be hard to model a >>> superior opponent. :-) >>> The winrate has to be somehow brought into a meaningful range, same idea >>> as with dynamic komi. >>> But instead of taking dynamic komi, the move generator could deteriorate >>> the opponents answers, until the winrate can make >>> good and bad moves distinguishable. >>> >>> Stefan >>> >>> * 20 minutes plus five 30 second byoyomi periods >>> _______________________________________________ >>> 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 >> >_______________________________________________ >Computer-go mailing list >[email protected] >http://dvandva.org/cgi-bin/mailman/listinfo/computer-go -- Hideki Kato <mailto:[email protected]> _______________________________________________ Computer-go mailing list [email protected] http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
