On Thu, May 12, 2011 at 14:25, Kahn Jonas <[email protected]> wrote: >> I have a question: what exactly is the horizon effect in the context of mc >> go? >> This paper seems to see it mainly as the problem, that threats will not >> always be properly answered in the playout. >> I thought that the horizon effect was about pushing an unwanted, but >> unavoidable result out of the tree by interposing ko threat type moves in >> the tree. > > The problem is that the horizon effect is felt also without ko threats: > let there be a lost semeai on the board. > The tree sees that it is lost, so it does not play there. No need for a > ko threat or anything, this is just a cold area, and should not be > played at (at least not until a ko fight or some special circumstance). > So that the semeai is solved during the playouts, with a random result.
This is a good example of horizon effect. Another one could be: There are 10 independent subgames on board. MCTS is reading them out. Each will be read out in some branch of the tree, but most of the branches will have some of them left unread. When playout leaves the tree, a static policy takes over. It can't learn from expirience so it will consistently misevaluate unread subgames leading to evaluation errors. I think that low exploration may work better than high exploration partially because of this problem. With low exploration at least on Principal variation most subgames will be read out. > Hence the evaluation function explodes... > > Jonas > _______________________________________________ > Computer-go mailing list > [email protected] > http://dvandva.org/cgi-bin/mailman/listinfo/computer-go > -- Łukasz _______________________________________________ Computer-go mailing list [email protected] http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
