On Oct 1, 2010, at 11:00 AM, Petr Baudis wrote: > > This is extremely interesting paper, thanks for making it available! > I was working towards something similar, but this is very valuable. > > I'd like to ask about couple of unclear things: > > * Did you train for all possible 3x3 pattern combinations? I didn't > have the time yet to check what 2100 base patterns sum up to. ;-)
Hi Petr, I don't understand this question. We learn one gamma for each 3x3 pattern. There is no such thing as "pattern combination". > > * More importantly - how do you handle tenuki? Does the playout > *always* play a local move if any is available, or do you include tenuki > with some value within the probability distribution? The same 3x3 patterns are used both for tenuki and local moves. Local moves are made more probable than tenuki thanks to the "contiguous" feature. > > * Your ko feature seems a bit strange. Did you find that including ko > solution by connecting the ko is not beneficial and it is better to > restrict it only to captures? Overally, does the inclusion of the ko > feature make a large impact on playing strength? > Aja will answer this. I expect the ko feature is not very important. Rémi _______________________________________________ Computer-go mailing list [email protected] http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
