2011/7/4 Andrés Domínguez <andres...@gmail.com> > 2011/7/4 Erik van der Werf <erikvanderw...@gmail.com>: > > > > If the program can play 30k level moves against you and still win its > > a strong indication that you were playing a hopeless position. Sure > > it's annoying, but that's because you should have resigned. > > I don't know if you are a Go player or not. If you are I suppose you > never lose by points, always by resign. If you are losing by 6 points > in the final stage of the small endgame you resing, isn't it? Some > players (like me) like to do the counting stage of the game, _even_ > if we are losing, and we don't want the other player to play stupid > moves to win by 0.5 points. If you are this kind of player, there is a > problem with the bot endgame. >
I think with this statement you have clearly identified the "disconnect." There are really two issues here: 1. How to make the program play these special positions like a human. 2. How this affects the playing strength. This has come up numerous times over the past several years. It's almost always in the context of point 2. The assertion is that this is that by "fixing" this the program should play much stronger. I don't think any of the really strong bot authors are saying this but I could be wrong. There is some hope however that there is a way to fix this that might make the program slightly stronger and some have claimed that against humans it would get a few "swindle wins" when on the losing side of this. I think this is probably true because the longer you make the human play the more likely it is that he will make some error. I'm not really interested in focusing my energy on winning like this although I'm sure it has some value. I think it remains to be demonstrated how to do this however without actually making your program weaker. Point 1 above is basically solved (although it can always be improved.) If you don't like that sort of behavior you can fix it without substantially weakening the program. I think Dave Fotland does not have too much trouble with this as his algorithm is a kind of hybrid of go knowledge from he older classical go playing algorithms and tree search. Imposing knowledge directly to the tree (so that it focuses on non-random looking moves even when dead lost) of course solves the program in general. I don't know any of the details of Davids program but I think that is the basic gist, perhaps David will correct me. > > With you way of thinking the bot should never resing, because it > reduce the probability of winning. Resign is only intended to finish > a boring game (human politeness), not to play better, not about > increasing the number of bot wins. Apart from that to humans is > not the same lose by 6 points or by 100 points, like is not the > same if a cyclist gets second in "le tour" by 3 minutes or by 2 > hours. > If you want to write a strong go program, the primary goal has to be to win the game. If you like the cycling analogy, winning by 2 hours is much more impressive, but if your "only" goal is to with the tour by 2 hours, you HAVE to sacrifice winning chances to do it because winning by 5 minutes is really unacceptable to you. This basic concept explains why counting points is so inferior to counting actual win statistics, and the demonstration of this helps us to appreciate the importance of trying to measure specifically what you are interested in. You can optimize it to win points or to win games. It's really hard to do both optimally and in fact may be impossible. You cannot have 2 competing goals and expect that you can do 100 percent justice to both. All you can do is find a compromise you are willing to accept. > > IMHO this thread is about a Go player who is annoyed with the bot > endgame. If the author of the bot don't mind about human players, > all is right, but Leon wants to collaborate to change the situation. > I mean no disrespect to Leon, but he seems to think this is a brand new issue that he discovered on his own. He has to first go through the process of thinking it should be easy to solve, thinking of naive solutions, finding out that they do not work, etc. Basically he has to gradually build up the sophistication of his understanding of the issues and he has started that process by joining our group. You have to walk before you run. So far nobody has proposed anything really fresh and new on this subject so who knows, maybe Leon will? > Maybe can make two modes, "polite" when humans want to play > or watch the game, "boring" when only matters wins. > Hasn't this already been solved? You can make the program look more natural without making it play much weaker right? Leon is a VERY strong player and I really hope this is NOT what he chooses to work on. Why would we want to take a potential talent and put him to work on something that is not that interesting and which will not have much impact on the strength of the program? Does Leon just want to polish the program up a bit or roll up his sleeves and work on interesting problems? Don > > Andrés > > P.D: I'm not sure the losing points moves of bots really increase > allways the real probability, maybe is an aliasing problem, a > statistical significance problem. > _______________________________________________ > Computer-go mailing list > Computer-go@dvandva.org > http://dvandva.org/cgi-bin/mailman/listinfo/computer-go >
_______________________________________________ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
