On Tue, Jul 5, 2011 at 10:01 AM, Leon Matoh <[email protected]>wrote:
> Hi, > > It started with endgame. > > It is a hack and i like it. > > When I normalized winrates that means I > assign a value to best move. > > When bot maximazes winrate it actually > plays the best move it can see in the tree. > When bot tries to kill and plays agressive > it tries to maximaze the result. > When plays strange moves when winning it is > because it cannot differentiate between winrates close to 1 > > When I say normalized winrate with random komi > I actually mean << alfa-beta prunning with statistic >>. > > Of course this "hack" will play a little worse now > because it was not designed for this even if > all things have their counterparts in alfa-beta prunning. > Please see for yourself. Or somebody else if there is > someone willing. And normalization takes a little > more time. > > Now if you suspence your disbelief and look at it from > different perspective. > > I know you think I am a little driven, but > there is nobody else to be on my side yet. > This is not a matter of being on your side or against you. There are two issues as I have pointed out: 1. Is the program now more cosmetically appealing? 2. How does the change affect the actual strength of the program. It's ok if the strength is impacted negatively if you can live with that. But you have not yet stated what it is you expect from this change. Do you expect this to make the program play stronger? Are you suggesting this as an improvement in style or in substance or both? There is no incorrect answer here. I suggest that you submit both versions for testing (post both version somewhere where we can all get to them.) If the difference in strength is major we will know with a few hundred or a thousand games. If the difference in strength is minor it will be relatively obvious after 1000 games. Let's say we test and see that it loses 30 ELO but makes the program much more appealing in style. That's not a bad thing but each user will have different sensibilities on whether this is worth it. I suggest that a good way to achieve what you want is to apply a first order score to all the moves at the root of the tree based on human based statistics. Normal and natural moves first. Apply some very slight incentive so that they are considered before other moves, all else being equal. Then when it doesn't matter the program will at least play moves that do not look so random and your sense of aesthetics will not be so offended. Don > > My goal is to play the perfect game. And this > is a step to that direction in my opinion. > I hate to tell you this, but it's probably a step in the wrong direction. You are not going to get a perfect game out of computer go programs in your lifetime, although you might get very nice and well played games. Don > > > Leon > > > ______________________________**_________________ > Computer-go mailing list > [email protected] > http://dvandva.org/cgi-bin/**mailman/listinfo/computer-go<http://dvandva.org/cgi-bin/mailman/listinfo/computer-go> >
_______________________________________________ Computer-go mailing list [email protected] http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
