Accelerated UCT does this: https://www.conftool.net/acg13/index.php/Hashimoto-Accelerated_UCT_and_Its_Application_to_Two-Player_Games-111.pdf?page=downloadPaper&filename=Hashimoto-Accelerated_UCT_and_Its_Application_to_Two-Player_Games-111.pdf&form_id=111&form_version=final
Rémi On 6 mars 2012, at 20:50, Stefan Kaitschick wrote: > On Tue, Mar 6, 2012 at 3:16 PM, Álvaro Begué <[email protected]> wrote: > I think the solution is the introduction of some flavor of minimax > into the tree search. For instance, once a node has been visited more > than a certain number of times, the score that we'll back out from it > is just the score of the best child, instead of an average that might > be polluted by bad moves. > > Of course someone else would have to figure out the details of how to > do this. :) > > > Álvaro. > > This isn't a go specific problem, is it. > But that means that either this is a wild goose chase, or there is room for > improvement of general MCTS here. > Winrate is certainly the legitimate criterion to distinguish between siblings. > But should the parent node really get full credit/blame for the weaker > continuation (in the case of go refutation) attempts? > This is really one side of an extreme, minimax being the other side. > Maybe something in between could be useful. > The weight of a node could grow with greater than linear speed in respect to > the number of attempts made. > > Stefan > _______________________________________________ > Computer-go mailing list > [email protected] > http://dvandva.org/cgi-bin/mailman/listinfo/computer-go _______________________________________________ Computer-go mailing list [email protected] http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
