On Tue, Mar 11, 2008 at 09:05:01AM +0100, Magnus Persson wrote: > Quoting Don Dailey <[EMAIL PROTECTED]>: >> >> When the child nodes are allocated, they are done all at once with >> this code - where cc is the number of fully legal child nodes: > > In valkyria3 I have "supernodes" that contains an array of "moveinfo" for > all possible moves. In the moveinfo I also store win/visits and end > position ownership statistics so my data structures are memory intensive. > As a consequence I expand each move individually, and my threshold seems to > be best at 7-10 visits in test against Gnugo. 40 visits could be possible > but at 100 there is a major loss in playing strength. > > Valkyria3 is also superselective using my implementation of mixing AMAF > with UCT as the mogo team recently described. The UCT constant is 0.01 > (outside of the square root). > > When it comes to parameters please remember that they may not have > independent effects on the playing strength. If one parameter is changed a > lot then the best value for other parameters may also change. And what > makes things worse is probably that best parameters change as a function of > the playouts. I believe that ideally the better the MC-eval is the more > selective one can expand the tree for example.
Typically, how many parameters do you have to tune ? Real or two-level ? If you consider a yet reasonable subset of parameters, an efficient way to estimate them is to use fractional factorial design for the linear part, and central composite design for quadratic part (once you know you are already in the right area). You are much more precise than with change-one-at-a-time strategies if there is no interaction between parameters, and you can detect interactions. Since anyhow computers are used, it might be possible to choose sequentially automatically new values of the parameter that optimize your efficiency. That's a very interesting problem, with much work on it in the statistical community, but I do not know that very well (neither the former designs, but those are easy). Alternatively, especially with a very high number of real parameters, derivatives of MC techniques can be efficient and easy to implement: particle filtering or swarm optimization in particular. Jonas _______________________________________________ computer-go mailing list [email protected] http://www.computer-go.org/mailman/listinfo/computer-go/
