Harri Salakoski wrote: >>> The *average* length of a 9x9 playout is roughly 100 moves. >>> The max length is much larger. >> The *average* length of a 9x9 playout is roughly 100 moves. >> The max length is much larger. > Hmm, sorry if this is old subject but does it effect much for playout > quality if I cut playouts for example max 110 moves in 9*9 board? Yes, that will significantly hurt your play-outs. Do you throw away the results or score it as is?
I have a very liberal cutoff on my program - from any position I play to move 81*4 in 9x9 as long as I have played at least 20 moves. > Is that studied subject for almost random playouts. What is commonly done is something called the "mercy" rule where you stop the play-outs early if one side appears to have an overwhelming advantage. Even this is somewhat risky if there is a large group with no eyes. I don't use that rule but it's been reported to be anywhere from no benefit to a minor benefit. I have not tested it, but the most you can hope for is a relatively small speedup for a small risk. > > Another thing, do you include random moves for playouts, after some > number of playouts or when there is "K" number empty points or using > some other way. This is all a black art - you must try many different things and see what works best for you. I haven't tried this, but it would be a major speedup to revert to "light move strategy" after a few heavy moves are tried in each play-out. My heavy play-outs are 3 - 4 times slower than the purely random play-outs. I suspect most of the benefit occurs in the first few moves. Is this what you are suggesting? > > By the wat made java board for random playouts it is currently 300000 > games /13sec 9*9 board having max 110 moves(double core 4000+), I have > no lightest clue how to make it faster as optimized it as much i can, > using two threads it is 7 seconds/300000 games. Have think that maybe > somekind of state machine could be faster. > > t. Harri > > > >> On a 2.2Ghz Athlon64, I get about 10 000 playouts/second, at an >> average of 100 moves per game this is 1 million updates/second. >> >> There are many programs that are much faster. On the same hardware >> libego would be about 6 million updates/second. >> >> 19 x 19 is a bit slower, because strings are bigger on average. >> >>> >>> This also explains that when I read the games MoGo against GNUGo, >>> toward >>> the >>> end of the game, GNUGo would play PASS, but MoGo would continue to >>> play at >>> some very uncommon positions that a normal player would never consider. >> >> Pass behaviour has little to do with the playouts themselves. >> >> -- >> GCP >> _______________________________________________ >> 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/ > _______________________________________________ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
