Magnus, could you recap your dynamic komi algorithm?
On Mon, Oct 4, 2010 at 3:52 PM, Don Dailey <[email protected]> wrote: > Magnus, > This is a really good experiment, I think it is well past due for someone > to try something like this and report the results. > I know that it's not possible to make the experimental setup perfect, but > it is important to establish some kind of baseline result to work from. > I hope you are able to play enough games to be statistically significant. > I have seen way too many papers on games of various kinds where some > algorithm is tried and they "prove" it is helpful with 100 games and the > score is 55% or 60% - that is just not enough to be statistically > convincing. > Don > > > On Mon, Oct 4, 2010 at 2:05 PM, <[email protected]> wrote: >> >> Hi Don! >> >> I just started an experiment! First some thought about what I believe is >> the factors involved. And then I describe the setup and some preliminary >> impressions and results. >> >> Quoting Don Dailey <[email protected]>: >> >>> 1. boardsize could be a factor. >> >> I would not expect it to be useful on 9x9, especially dynamic komi would >> get incredibly unstable on small board towards the end of the game. So what >> I am testing is 19x19. >> >>> 2. it may be more feasible in handicap games. >> >> My previous experiments for even game showed little or no gain. My >> subjective feeling is that it is in high handicap games on 19x19 where >> effect can be strong. >> >>> 3. it may work better against humans for psychological reasons. >> >> I cannot measure that. >> >>> 4. Maybe I did not stumble on the right approach when I did my >>> experiments. >> >> With dynamic komi the MC evaluation suffers more from noise because when >> the win rates get close to 100% it even makes MC programs play randomly (or >> systematically from the top left corner or something similar). This problem >> becomes greater the stronger the program is. It is probably not so bad >> because a stronger program will also have larger difference in win rates for >> strong and weak moves. This is my explanation why Valkyria probably only >> benefits from. >> >> The other reason why dynamic komi may be good is that a MCTS program >> prefers high win rates over a safe winning margin leading to the 0.5 points >> syndrome. On 9x9 boards present programs do not appear to suffer from this. >> They are so strong that with a "safe" 0.5 lead they win without mistakes. >> And if they do suffer from it still it is not clearly visible. It looks like >> a middle game mistake and then they realize they are behind a wild fight >> occur. >> >> On 19x19 this is probably a major problem for most program without dynamic >> komi or some other greedy search method. If a game is long and the score >> drops to 0.5 long before the game is finished the program must play the rest >> of the game as strongly as the opponent. This is a disaster for black in >> high handicap games. In the worst case even an opponent just slightly >> stronger can win with a large handicap, because the game become even too >> soon. >> >> Ok, enough babbling about the stuff dynamic komi could possible fix. >> >> I just started an experiment to make Don happy. It is not so easy to make >> such an experiment completely satisfactorily but it will be a first step to >> test Valkyria at least. >> >> I want to show that Valkyria plays stronger as black with 9 handicap >> stones on 19x19. The problem is that I would need an extremely strong >> opponent to beat Valkyria as white under normal time constraint. And I have >> no access to such an opponent. >> >> So I let Valkyria play Black with about 1s per move and White with 10s per >> move approximately (1000 and 10000 playouts respectively). >> >> I run two conditions in parallel one where black use dynamic komi and one >> where it play without it. My hope is that the modest difference in playing >> strength will be enough to make a difference. In the worst case black wins >> 100% of the games in both conditions. >> >> The good news is that I think the effect of dynamic komi in this >> particular experiment is huge just from watching the very first games. >> >> With dynamic Komi black is really aggressive. It is tactically weak but >> fights in a way that the white groups are eyeless and have no territory. It >> kills large white groups in almost every game. >> >> Without dynamic komi white is allowed to establish relatively strong >> groups all over the board without a fight and gets closer to an even game >> for almost every move played. So far white won 5 out 7 games which I did not >> expect. 10k vs 1k should not be more than a 3-4 stones I thought. >> >> White has won 1 game out of 6 against black with dynamic komi. >> >> It is too early to make any conclusions yet but if the effect is as large >> at it seems to be I can probably give some convincing numbers already >> tomorrow. >> >> Caution is important also. As in all self play experiments effect tend to >> be exaggerated. The next step is this seems to be a good setup is to test >> against Fuego. I would also like to scale it up using 10k playouts for black >> and 100k playouts for white to if effect stays with better play for both >> colors. >> >> It may also be that the effect is double. White with h9 has winrates below >> 20% in these games and play desperate move rather than patient moves quite >> often. Maybe white can use dynamic komi too to some degree? Or perhaps a >> linear komi to make it play the opening more sound and then get more >> desperate later? >> >> -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
