Hello,
> Well, at least I learn at lot. E.g. that you are the author of MoGo :-) One of the authors, but yes :). > As an old chess-programmer the unit is not games/second but nodes/sec. > Making a move and undoing it (if undo is done at all). Thats the basic unit > in any game. Ok, but as for one node (a position) we make one simulation (a complete game till the end), then nodes/sec == simulations/sec == games/sec. Indeed, when from a node (in the tree) you make one move you create a new node (a child). So at every simulation you create a new node. > In your MoGo paper you mention the First-Play-Urgency FPU. I did not > understand precisly what FPUs is: > Is it: If a node is visited first (or not all moves have been tried out) > the nodes are not selected with uniform probablity but e.g. Capture and > Atari moves have a higher probablity? Sorry about that. I received other emails saying it is not clear, so it must be very unclear. As I have no time to modifiy the report right now, here is an explanation: First Play Urgency is not related to the MC simulation part, but to the tree search part. UCT says that, on a position, you have to try all the moves once before using the UCB formula. FPU is here to avoid trying all the moves first. You simply extend the UCB formula for unexplored moves, saying that if there is no yet a child for this move, you allotate to it the FPU. Then, if FPU is big (say 10000), then you have exactly UCT (because 10000 is bigger than all the X_i+sqrt(log(n)/n_i) for the explored moves). If FPU is 0, then you will explore the first move, and then only this one, because X_i+sqrt(log(n)/n_i) is always greater than 0. Now if you set a FPU around 1, asymptotically you will explore all moves at least once, but you can explore a move twice (or more) before exploring another move. Than allows you to go deeper for the deep nodes (not so much explored nodes). I hope it is clearer now? Sylvain _______________________________________________ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
