On Thu, Jan 24, 2008 at 03:19:52PM +0100, Magnus Persson wrote: > > As a rule of thumb I want 300 games for each datapoint and at least > 500 if I am going to make any conclusions.
Ok, I think we start to have those 500 games. To my eyes, FatMan shows a clear turn in the curve at FatMan_03. Before that point the curveis practically a straight line, and after that it is pretty well linear too, but with a considerably lower scope. For MoGo, there may or may not be such transition around MoGo_04. I still find it an odd coincidence that these turns happen around 1400-1600 ELO, which is close to the limit of what a pure MC-program can reach. So, I have two questions for the list. 1) I am not much of a statistician, so I can still not judge how much I could trust this thing I see in the curves. 2) Assuming there is something in it, what can we conclude from it? That having a better evaluation function helps an UCT program, but more on lower levels of play, where its effects are greater. That there can be changes in the scalability of UCT programs. We (may) have identified one. Will there be another lurking somewhere? - Heikki -- 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/
