Your dynamic komi results are very convincing.. But shouldn't we just concentrate on maximizing score instead of winning rate in the beginning of the game?
Best Regards, Lukasz Lew On Tue, Oct 5, 2010 at 14:01, <[email protected]> wrote: > Here are the results from my overnight experiment. > > Black had handicap 9 and played 1k playouts per move (about 1 second per > move) > White played 10k playouts per move (about 10 seconds per move). > > Both players resigned when Win Rate became less than 1% at root > > 50 games were played. > > Without Dynamic Komi: > > Black Wins 16% +-5.2 > Game length: 220.4 +-6.1 > > With Dynamic komi > Black Wins: 70% +-6.5 > Game Length: 180.1 +- 10.1 > > Standard Errors are those as given by twogtp which I used to run the > experiment. > > As you can see no statistics is needed other than descriptive. The effect is > in the range of several hundred Elos. Or perhaps 5 stones or so if I > modified the handicap to titrate the difference. > > I also included game length to show that Black forces white to resign early > with dynamic komi. The shortest game was only 84 move. In the position all > white stones were dead or cut of in eyeless groups with no weak groups for > black. One group could perhaps live but it would be a small group with no > chance of further territory. > > This effect is huge. But this is for selfplay and I think I now understand > what happens. The change in playing style is dramatic for black. You can see > for your self since I attached some of the games in a zip. Email me if you > want to see them all. What happens is that black relentlessly attacks white > groups. It has a very clear direction of play. Without dynamic komi it often > plays vague moves elsewhere. With dynamic komi it sort of uses sente. > Without it does not care. > > What does not happen is that black *does not play better shape*. I think I > long time ago criticized dynamic komi proponents because they seemed to > believe dynamic komi would make the program over all stronger in all aspects > and I instinctively knew that adding go knowledge to playouts and tree > priors is more important. Black with dynamic komi play really weird moves > (1s of thinking time) but every weird move seems to at least force white to > defend or counterattack leading to endless fighting and gives black overall > strategic initiative. So the effect of dynamic komi is all about strategic > evaluation, but has very little effect on the reading ability. > > The self play factor is important because white who plays without > dynamic komi expect black to play aimless moves and thus white thinks it can > get away with moves that only works if black plays elsewhere. So dynamic > komi become very effective for black in this case, because white expect a > different style of play and does not adapt. > > So my conclusion is that I also need to do something about white. Easy would > be to enable dynamic komi also for the player who is behind. But it feels > wrong to do so. Perhaps a quickly shrinking linear komi for white makes it > play steady in the opening and then later it can play in full desperation > mode. > > I also think I can be very satisfied with my implementation of the dynamic > komi. Although the effect is certainly smaller against strong humans it is a > big change. > > I am now going to repeat this experiment with 10k vs 100k playouts which > would then test black as it would play in actual games with normal time > constraints. > > -Magnus > > _______________________________________________ > Computer-go mailing list > [email protected] > http://dvandva.org/cgi-bin/mailman/listinfo/computer-go > -- Łukasz _______________________________________________ Computer-go mailing list [email protected] http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
