Looking at it from a distance, this sounds like a fancy way of saying that you created an opening book. This may sound a little strange and a mis-characterization of your effort, but please entertain the thought for a while. What you are attempting to do is to bias the move selection in the opening phase using priors on the top-30 moves. Perhaps Fuego's opening book code would allow you to import your weights outside of the tree/playout code. Now, typically opening books contains well-defined lines of play, whereas yours would be a model, so integration may not be that straightforward. You would also lose the guidance inside the random playouts.
Rene PS. Welcome to the list. On Mon, Jun 24, 2013 at 8:33 AM, David Briemann <[email protected]> wrote: > Well it is an attempt to improve the playing strength, but that won't mean > that it succeeds. > > What I do is the following(in short): > I have a trained move predictor model which consumes a board situation and > outputs beliefs for every playable move. > I want to use it to bias the search tree for the first N moves of a game > (opening phase). > > So when tree search generates all legal moves, the predictor will score > them and only consider the best X move as legal moves. > > It then should be forced to play "good" opening moves(of couse only if the > predictions make sense). > > David > > > 2013/6/24 Don Dailey <[email protected]> > >> >> On Mon, Jun 24, 2013 at 7:58 AM, David Briemann <[email protected]>wrote: >> >>> I'm beginning to think that I didn't understand the tree search part >>> correctly. You say the tree search generates moves too. I thought moves >>> were only generated in playouts and the tree search part was to follow >>> already played lines until it reaches a position which has not been played >>> out. Probably that's the location were I have too look into. >>> >> >> I don't know the gory details of the implementation, but clearly the >> tree portion of the search considers all moves (sooner or later) and much >> has been written about how MCTS is admissible - at least in theory. That >> means it would, if given enough time and memory, play perfect go and will >> consider every legal move at some point. But we know that playouts are >> not fully random and in many positions will only play a limited number of >> moves (perhaps just one) such as when defending atari. So the search >> tree portion is not constrained by only what the next playout move will >> return. >> >> Read the code - and perhaps any documentation that comes with this >> program. One this is clear to me though, if you impose patterns >> non-probabilistically on the tree you will weaken the program considerably. >> The reason MCTS works so incredibly well is that we have put patterns >> in their proper place, as move guidance and not as a plausible move >> generator only. The old style weak programs were heavily pattern based. >> So I may be misunderstanding what you are trying to do - is this a >> study of some kind or a real attempt to improve the program? >> >> Don >> >> >> _______________________________________________ >> 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 >
_______________________________________________ Computer-go mailing list [email protected] http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
