On Wed, Nov 25, 2009 at 10:44 AM, Heikki Levanto <hei...@lsd.dk> wrote:
> On Wed, Nov 25, 2009 at 09:01:22AM -0500, Don Dailey wrote: > > You could of course just play games where you choose each player > randomly. > > If you have 256 feature you have a ridiculous number of combinations, > more > > than you could possibly test but before each test game you just pick a > > combination of features randomly for each player. In this way about > 1/4 > > of the games will be relevant for each feature, regardless of how many > > features you are testing. (1/2 will have the feature on and half of > those > > games will be against opponents who have the feature off.) > > Wouldn't it be more effective to choose one player randomly, and make the > other a "mirror image" of it? That way, every game would give some > information of the relevance of one setting. At least in the very > beginning... > That would not be effective at all. With 256 features you are (for all practical purposes) never going to see that exact combination of features again. In very general terms, you are probably going to be more interested in how a few terms interact than how many interact. Of course this method only tries to understand each feature independently, but I think this has some validity. I don't claim it will cover all the bases however but it might be a good place to start. Of course there is nothing that prevents you from observing from the data the interaction of any specific combination of features, but the amount of data you will get for specific combinations of feature is going to be drastically reduced with the number of features you wish to look at. I will assert that looking at each feature independently is half the battle, and looking at combinations of 2 features is going to cover a higher percentage of the remaining cases, and so on. And if you find interesting interactions, you can COMBINE them in a separate test - by basically considering feature x and y together as a single compound feature. You would only do this when you have a high degree of confidence that these 2 features definitely have some kind of synergy. You might even do this with features that have a negative interaction, perhaps taking care that they are never tested together. - Don > > -- > Heikki Levanto "In Murphy We Turst" heikki (at) lsd (dot) dk > > _______________________________________________ > computer-go mailing list > computer-go@computer-go.org > http://www.computer-go.org/mailman/listinfo/computer-go/ >
_______________________________________________ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/