When bad playouts recommend bad moves to the tree, it isn't quite as bad as it seems because the fact that we focus on these moves helps clarify them and we see them for what they are.
Of course that is not an argument against using high quality playouts. I believe there is clearly a time/quality trade-off here. The playouts have a very strong influence on the how the tree gets shaped and expanded. It has been called a "black art" because the best moves do not always improve the program and that's probably because it's more about the shape of the tree than it is the absolute goodness of the moves in the playouts. So I personally don't believe we can consider the playouts separately from the tree, as if all you have to do is figure out a good way to do playouts and they are plug-in compatible with every program. I'll bet they are not. - Don On Thu, 2008-10-16 at 09:22 +0900, Darren Cook wrote: > >> http://dcook.org/compgo/article_the_problem_with_random_playouts.html > > I admit not reading this thread carefully enough to understand what the > > argument is about. So I just contribute with following statement: > > > > Positions like this are very easy to handle for MCTS based programs that > > uses heavy playouts. > > Thanks Terry for making the SGF. > > My point with the position is that random playouts handle this position > poorly, because though white is ahead his position is more fragile. I.e. > when black attacks either corner white usually needs to reply. But, > being random, and with more tenuki moves than response moves to choose > from, it will choose to tenuki. More times than not some white stones > will die and so the position ends up being reported as good for black. > > Of course MCTS, within a few hundred playouts, will discover white > shouldn't tenuki . But when this position is found 20 ply deep in the > tree it will only receive a few playouts, so bad information is being > passed up the tree. The point of the whole article was that these > unbalanced positions are rare in games between weak players but dominate > games between strong players (i.e. play is on the knife's edge). > > I see Gunnar pointed out a similar conclusion elsewhere in this thread > (quoted below). > > For the given position very simple move-weighting rules such as: "when > the last move ataried something, try and capture one of its neighbouring > chains" will cope with two out of black's three attacks in the > bottom-right, and a similar proportion in the upper-left. Even just > weighting moves near the opponent's last move would be enough here I think. > > So, the interesting (and difficult) question is: how heavy is good enough? > > Darren > > Gunnar wrote: > "That's why uniform simulations are badly biased; solidly connected > stones almost always win against more loosely connected but perfectly > sound formations, causing the program to play very heavy and > inefficiently." > >
signature.asc
Description: This is a digitally signed message part
_______________________________________________ computer-go mailing list [email protected] http://www.computer-go.org/mailman/listinfo/computer-go/
