On Wed, 2008-12-31 at 00:25 +0100, Rémi Coulom wrote: > Don Dailey wrote: > > You are right, the d3p version rallied to come from behind and staged > > an exciting and dramatic comeback: > > > > Rank Name Elo + - games score oppo. draws > > 1 d3p 2016 77 75 21 52% 2000 0% > > 2 base 2000 75 77 21 48% 2016 0% > > > > > > - Don > > If you'd like to try a simple pruning scheme that improves playing > strength on 19x19, then I'd suggest progressive widening. It only works > in the tree, not in the playouts. You don't need complex patterns for > progressive widening to work. You can simply use distance to the > previous move. Search moves at distance 1 from the previous move for N > playouts. Then add moves at distance 2 for N*x playouts. Moves at > distance 3 for N*x*x. N and x are the two parameters that need tuning. I > expect you can get tremendous strength improvement on 19x19 with this > simple scheme.
I'm not sure I understand - when you say N playouts, do you mean N visits of that node? Because once you visit a node, you expand it, no longer doing playouts from that point. For instance if e5 is played (from the root position) are you saying we would only look at the moves touching e5 the first few times e5 was visited, then start looking at distance 2 for a while, and so on stopping after 3? The reference bot of course does not build a tree, what I'm actually looking for is a way to produce a medium strength but really simple bot that does not build a tree and just has a lot of playout magic. I wonder if this behavior can be emulated in the playouts somehow? It's not so simple because we are not expanding a tree, and when we look at some moves more often than others we get more statistical noise in the moves we don't look at, and this can make them look artificially good or bad. I can definitely see how such a scheme would work well in the tree. - Don > > Rémi _______________________________________________ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/