On Sun, Jan 23, 2011 at 11:00 PM, Darren Cook <dar...@dcook.org> wrote:
> > On Fri, Jan 21, 2011 at 08:31:04PM +0900, Darren Cook wrote: > >> Dynamic komi. (E.g. if it thinks it has only 30% chance of winning at > >> 7.5pt komi, but if you reduce the komi to 5.5pts it thinks it has a 55% > >> chance of winning, then reduce the komi to 5.5pts: it will play an > >> intelligent looking endgame and lose by 1.5pts.) > >> > >> This will make it weaker overall (because it won't try so hard to cause > >> trouble) > > > > What do you base this statement on? It's rather controversial for me. :-) > > Without dynamic komi the program will choose the move to maximize its > winning rate. If dynamic komi causes a different move to be chosen then > it implies it is choosing a move that it thinks has a less than or equal > chance of winning. > > Also, I believe Don did some (self-play?) experiments a few years ago > and the dynamic komi version lost more; I think this is where Don's > coolness to dynamic komi comes from (apologies if my memory is inaccurate). > My coolness is based on experiments with thousands of games, or even tens of thousands of games. The only way you can reliably measure a small improvement is to play many thousands of games. If the improvement is large, you might get by with a few hundred. If your program is well developed no single improvement is going to give you more than a few ELO points of improvement so you are pretty much required to play several thousand games and I doubt many here are doing that. If you use a tool such as bayeselo you will get error margins which should prove to you that a few hundred games is not enough to measure small improvements (say less than 10 ELO.) I was not able to find an implementation that was within 10 ELO of the non-dynamic version. I know a lot of computer chess developers who accept ideas based on 100 test games or a "good feeling" or even watching it play a couple of games and announcing that they can tell it's playing much better. Nevertheless, that does not mean I stumbled on the right implementation and proved it does not work. It's quite possible that I just never found a good way to do it. > However, Magnus has some experiments with Valkyria where the dynamic > komi version was stronger even in non-handicap self-play games. > > >> I.e. program endgame is generally stronger than the humans of the same > >> rank; chances are a 1-dan human will make a few 1pt or 2pt errors during > >> the endgame. > > > > I also think this is not obviously true at all. My observations have > > been that MCTS does not perform too well at all in very close endgames. > > (Though it is not a big disadvantage in practice since it is in the > > nature of MCTS to strive for deciding the game ASAP, i.e. in the middle > > game.) > > My study has mostly been of 9x9 games, and as long as there is not a > seki on the board the MCTS programs will practically never lose if they > have a winning position at move 30. (Extra condition for that statement: > Chinese rules.) > My experiments were are 9x9 too. I believe what was happening with my implementations of this is that it worked well most of the time, but not when it really mattered. When it didn't work, it was turning a simple win into a struggle and sometimes a loss. Don > > Darren > > > -- > Darren Cook, Software Researcher/Developer > > http://dcook.org/work/ (About me and my work) > http://dcook.org/blogs.html (My blogs and articles) > _______________________________________________ > 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
