What's your general approach? My understanding from your previous posts is that it's something like:
UCT search using Silver's beta formula and UCB1 with Win-rate and Rave for choosing a child (I use basic UCT with win-rate and Rave, and the original MOGO beta formula). UCT search is biased with variable sized patterns extracted from strong games using something similar to Remi's algorithm. I don't know if you bias the initial rave values, or bias the initial win rate, or add a third heuristic term to the UCT formula. (I bias the rave value, initially with a simple heuristic, and later with Many Faces' knowledge). I don't know how you accumulate the rave value. (I just use all moves that lead to a win for either side, in the whole game). You apply UCT to every legal move. (I apply UCT to a pool of moves that starts with one move and gradually increases to 30 moves. I use rave to choose which moves to promote to the pool). You have powerful playouts with tactical search in the playouts. How many playouts per second do you do per CPU on 19x19? (I have fairly simple playouts - local responses for 1 and 2 liberty neighbors, without local search, 3x3 local patterns, don't fill eyes or seki or make stupid self Atari.) Does this sound about right? > > Your implementation must be very different from mine. Actually I don't > use Progressive widening (or unpruning) at all. It's a mystery to me why > others say it does work. > > -- > Yamato > _______________________________________________ > computer-go mailing list > computer-go@computer-go.org > http://www.computer-go.org/mailman/listinfo/computer-go/ _______________________________________________ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
