Dynamic komi and some other tricks work quite well. Thanks to Ingo for pushing dynamic komi until I figured out how to make it work well. Often the playout have some bias due to a misread in a fight, so it's important for the bot to keep its lead.
If you look at kgs games with strong bots, 0.5 wins are now very rare. Example: manyfaces: http://www.gokgs.com/gameArchives.jsp?user=manyfaces David From: [email protected] [mailto:[email protected]] On Behalf Of Don Dailey Sent: Saturday, June 08, 2013 12:29 PM To: [email protected] Subject: Re: [Computer-go] Narrow wins Stefan, To a bot this always comes down to risk analysis and it's no different than what we do. In fact you just did your own risk analysis here of the computers play and have decided that it does risk analysis wrong so your thinking is no different from the bots. To me the biggest issue isn't winning by 0.5 but not fighting when it's losing. I have a feeling that most of this is not about improving the play of the bot although it's always cast that way, but in reality it's about not offending our own sensibilities. We just hate to see a bot win by 0.5 when it probably could have won the entire board or most of it. And of course not fighting when losing is not much fun to watch when it may still have a realistic chance of winning due to an opponents mistake. The standard proposed solution is komi-manipulation in one way or another. But what we are probably really after is something to do with "opponent modeling", giving the computer more of a fighting spirit instead of playing overly logical. I sense there must be a better solution as komi-manipulation just seems to me to be really illogical. It's like fixing the books because we don't like the numbers. I have failed to keep up much with computer go since getting involved with my chess program Komodo. But I see that this problem is still being talked about. Has there been any progress on this front? If this were computer chess we would modify the evaluation function, but the evaluation function in GO is really based on the play-outs. At the end of every playout we have a final board positions that returns some information. Some positions will return bigger "wins" than others (or smaller losses) but we know scoring this way is terrible. But maybe that information can be effectively used to shape the tree towards positions that win bigger? Sometimes it's a matter of which node to explore first - so it is quite likely that at least in some cases the computer does not make any distinction between a big or small win and thus we could possibly push it in that direction in the tree part? Maybe that is naive, I don't know. So what is the state of the art on this now? Don On Sat, Jun 8, 2013 at 2:21 PM, Stefan Kaitschick <[email protected]> wrote: Humans may be predisposed to the fallacy of greed. But bots have there own fallacies. Happily winning by 0.5, when a higher win at almost the same perceived risk is available, is a kind of full knowledge fallacy. A bot can lose an easily won game by merrily giving away everything but the last half point, then have the last point stolen from him by a single bot misread. So if you feel that the bot you're playing against is misjudging a particular situation, don't spring in on him unless that will put you in the lead. Otherwise, it's better to let him hand over other points that you need first. The bot might fix his problem, but that risk is preferable to having the bot go into business again when it is still ahead. Stefan _______________________________________________ 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
