> Hi, > > I'm also interested at the same thing. Glad I'm not alone. ;)
> > It sounds good but it seems to lack the check of whether a given move is > > first played in a given intersection. When you add that, it because a > little > > more tricky to update in the tree. I see, I missed that. I actually prune second (third, fourth...) moves at the same intersection in the playouts. However, this means that I can't just count every 2nd move as a move for the given node color, so I'd have to change the update loop. >>You also update the value of each > move > > independently of the color, i.e. a position in which it is black turn > will > > be update with white moves. You should restrict the update. This I don't really understand. Are you talking about the issue raised by basically doing j+=2 in the loop instead of checking whether the move color is the same as the node color? If yes, I understand. > I have a question about the the part were the stats are updated. > (l.15-25). having an array of amaf-values in every node seems very memory > intensive and I don't get how you would access these values. > > Something like searching for all nodes below the visited node with the > same > pos+color as the ones visited and updating amaf-stats directly in these > nodes > would make more sense to me. (but also costs more cpu time) I have been thinking *exactly* the same thing. When you keep the values in the node, it's easiest to keep boardsize*boardsize values to be able to access AMAF values directly by the childrens' positions. This wastes a bit of memory, but not too much. Keeping the value in the children is more cpu expensive but saves memory. For me, the first way seems easier to implement. How would you exactly do "searching for all nodes below the visited node with the same pos+color as the ones visited"? I think by the "visited node" you mean the first move played after root, is that correct? On a side note, I found that using plain AMAF for MCTS (so a single AMAF table for all nodes in the tree) actually made my program play about 100 elo *worse*! Clearly, it's necessary to implement proper RAVE to get a benefit. Thanks again, Isaac -- Sensationsangebot verlängert: GMX FreeDSL - Telefonanschluss + DSL für nur 16,37 Euro/mtl.!* http://dsl.gmx.de/?ac=OM.AD.PD003K1308T4569a _______________________________________________ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
