Are we sure that dynamic komi helps, or does it just make the program look a lot better? I know that a number of us tried this (including myself) on 9x9 and in every case it made the program play weaker. It's good news if it can be made to work. When I tested various forms of this I was very methodical about my testing - I require thousands of games and low error margins before drawing conclusions and it was always weaker and not by tiny amounts. I was ready to accept a slight weakening for more pleasing play but I always got more than a slight weakening.
I'm not trying to start an argument here - I know this goes around every once in a while. There were arguments about this but there was at one point a consensus by the authors who actually tried it, that they could not make it play stronger. The ones who disagreed were not authors or did not try it, they just argued it. There were a couple of experiments that were far from scientific which involved manually changing parameters and pathetically small samples. I would like to see it work and want to be wrong. I have some sense that it should be possible to make this work if done right. What is the state of the art on this? I am willing to accept any conclusion that is backed up by a statistically significant number of games (please don't give puny 200 game sample which tends to generate error margins of 50 ELO or more.) I just don't want us to draw conclusions based on someones opinion that they like how it plays now. I have no doubt that dynamic komi does make the programs look more natural and pleasing, but I still question whether it actually plays stronger. I have some ideas here: 1. boardsize could be a factor. 2. it may be more feasible in handicap games. 3. it may work better against humans for psychological reasons. 4. Maybe I did not stumble on the right approach when I did my experiments. We don't need to reiterate the arguments for why it should or should not work. I don't really want to get into the philosophy of this or opinions (including my own), I just want to know what have been found to actually work and how this was proved and some details on the empirical proof involved if anyone is willing to share this. Don On Mon, Oct 4, 2010 at 9:31 AM, Stefan Kaitschick < [email protected]> wrote: > With the right level of komi, dyn. komi is almost certainly beneficial. > One way to look at it, is that the komi holds some extra information about > the position. > The main problem ofcourse, is choosing the appropriate komi level. > I think a linearly decreasing scheme and a constant winrate scheme are both > somewhat unsatisfactory. > The point is that you want as big a point advantage as possible, but you > don't want to take a huge extra risk to achieve this. > I think that the optimal komi for a certain position should be searched for > by making several searches at different komi levels. > Because the goal is to find a winrate to komi profile, rather than actually > finding the best move, the opponents time could be used for this. > > Stefan > > > Quoting Darren Cook <[email protected]>: >> >> And I think Magnus (Valkyria) got a similar result (though not sure if >>> that was just in handicap games). >>> >> >> I got a more greedy form of dynamic komi. In even games against very weak >> players it plays aggressively until it has captured all stones (given the >> opponent play that weak). I have tested this on even games only in self play >> and it does not seems to change anything. But I like the style of play from >> handicap games and that is enough reason to use it even if I have no >> evidence of it playing better. >> >> The hard part of dynamic komi is that small changes to the komi has a >> large effect of the evaluation in the late endgame. With linear komi search >> is much more robust but the games do not get as much drama. >> >> Magnus >> _______________________________________________ >> 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
