I put up my old simple MC program on CGOS 9x9 as a reference bot. It is called Ike and IkeJr, Ike does 10000 playouts and IkeJr does 1000.
Here is how the play-outs work: 1. play uniformly random simulations. 2. Eye rule as described. 3. Play until no non-eye moves left, then pass. 4. 2 passes stop. After N simulations a move is selected using most-moves-as-first. I look at 3/4 of (5 + the remaining moves) but not going past the end of the game of course. I only consider moves for the side to move of course. I also DO NOT check to see if the move was already played by either side - so a move could get scored twice in 1 play-out or it could get scored even though the opponent played it first. (In experiments it didn't seem to make a different in the strength - but there is an argument for only scoring moves if the side to move in the game played it first.) So the move selected is the one that happened to win the most when played by the side to move accounting for komi. I believe this is enough to perfectly duplicate my algorithm as a reference point. Here are the bayes rated results on ELO (after less than 100 games.) Name ELO ------ ----- Ike 1291 (10,000 play-outs) IkeJr 1167 ( 1,000 play-outs) Note that on 9x9 you don't get much improvement after about 1000 simulations using all-moves-as-first. So there is some upper bound on strength here that you won't go beyond. There are a few simple things you can do to add perhaps 200 ELO to this. A few years ago when I did this, I tried not using all-moves-as-first to see if I could extend the upper bound on how strong this can be but I could not measure an increase. - Don On Tue, 2008-10-07 at 17:33 +0200, Denis fidaali wrote: > Hi, i have been searching a bit through the archive, and was not able to get > a complete list of the characteristics value for the light simulations. > > By light simulation i mean : Playing randomly the valid moves (No suicide > allowed), but not in the "true" eyes. I have a rule for defining a true eye : > all vertical and horizontal point around are the player's color. And > depending on the position : > - no edge around : NOT ( two diagonal or more are the opponent's color). > - one or two edge around : NOT ( one diagonal or more is the opponent's > color) > > The engine is written in java, and run on a quad core Q9300 @ 2.50 Ghz. > The code has been lightly optimized, and use pseudo-liberties to detect > captures. > > Here is what i get : > --------------------------- > > mean score =2.086382709065067 (black win on average by 2.08 points) > 78454.10819786233 Playout/sec (Trying to take advantage of the 4 cores) > Time=12.752397841 (in seconds) > Number of playout=1000478.0 > Mean moves per sim 111.05742754963127 (Number of move per simulation, > considering pass as a move) > Mean moves(no pass) per sim 107.37764148736903 (Number of move per > simulation, not counting passes). > > Note : as there are no seki possible in the simulation, the mean score for > the light simulations maybe bugged for seki, and still get the correct value. > _________________________________________________________________ > Installez gratuitement les 20 émôticones Windows Live Messenger les plus fous > ! Cliquez ici ! > http://www.emoticones-messenger.fr/_______________________________________________ > computer-go mailing list > computer-go@computer-go.org > http://www.computer-go.org/mailman/listinfo/computer-go/
signature.asc
Description: This is a digitally signed message part
_______________________________________________ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
