> > > > When I try AMAF (including RAVE) my program always wants to make early > > moves on the first and second lines. I've come to look at it as a > > hallmark of the algorithm. > > > > To combat this, I use a table with a move number for each position on > > the board. The program is not allowed to put a stone on that position > > before that many moves have passed-Unless there is already another stone > > nearby. > Dave, > Does this really help or does it just make the games prettier without > hurting the program??? I tried using tables like yours and could not > find an improvement although I realize the devil is in the details. > The "technology" I've used in these real simple programs is to > discourage but not prevent certain moves.? So if I tried this again it > would probably be in the form of how to apply incentives (or > discouragements.) > A nice way I discovered to do this is to gift some wins or losses for > specific classes of moves.??? I give captures a little incentive - so a > program is slightly more likely to play a capture than the simulation > specifies.?? The incentive increases with the number of simulations but > not linearly. > None of this applies to UCI or more sophisticated programs - I am only > considering the simple class of monte/carlo programs. > - - Don In AntIgo's heavy playouts, the table only comes into play for what would otherwise have been a uniform random move. The other heavy playout rules all involve moves that are close to other stones anyway. I think it helps a little but not a great deal when I factor in the extra computation. (And the playouts do look much prettier :)) Using the table for moves at the root is a big help, especially for low numbers of playouts. If I test AMAF as a way of selecting a move, with no UCT and no tree, then the table is the only way that I can prevent the bot from making idiot opening moves on the first line. How do other people do it? For my full-up program, the table is one of many features feeding into a neural net that controls progressive widening. In this case, the table means that UCT starts by concentrating on moves that are on a reasonable part of the board, but if the node is visited often enough, all positions on the board may be explored. - Dave Hillis ________________________________________________________________________ Check Out the new free AIM(R) Mail -- Unlimited storage and industry-leading spam and email virus protection.
_______________________________________________ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
