On Thu, 2007-02-08 at 08:59 -0500, Chris Fant wrote: > I thought that the memory boundedness was completely fixed by not > expanding a UCT node until it has been visited X number of times. > Just increase X until you are no longer memory bound. I don't recall > anyone reporting a loss in playing strength by doing this.
There is a loss of playing strength as X grows, but the loss grows very slowly. Obviously, if you set X really high, it would not build much of a tree and it would play very weak. I have been planning to make this dynamic in Lazarus, but I have not done so yet. The idea is that X starts at 1 but when the table fills up there is a consolidation pass where the table is rebuilt, child nodes may fold back into their parents and X gets reset to a higher value. As the search proceeds this may happen a few time, perhaps doubling the value of X each time. In this way the program continues to scale smoothly. There is never a sudden out of memory wall and you can set the table initiallly to any size that suits you. I haven't bothered because I cannot detect a loss of strength in values up to 100 but it's on my todo list - I know there is a small loss of strength even if I cannot measure it. Right now, Lazarus simply stops searching and returns a move when this happens, but I have X set high enough that it can think for many minutes. - Don > > On 2/8/07, Don Dailey <[EMAIL PROTECTED]> wrote: > > I think there are 15 first moves in 9x9 go if you factor out the > > symetries. > > UCT isn't good at evauating all the moves, it will pick one of them and > > spend most of it's time on it. But you could search each 1 at a time. > > > > The UCT programs are memory bound, so you could search each of these 15 > > moves 1 at a time and study the scores. > > > > - Don > > > > > > On Thu, 2007-02-08 at 04:41 -0800, steve uurtamo wrote: > > > > The average score can contain a very large proportion of losees if it is > > > > compensated by bigger wins. > > > > > > yes, it is easy to see how this might cripple the play of an MC player. > > > > > > that 90% territory win that requires 3 opponent blunders is tempting > > > enough > > > to ignore the fact that all other non-blundering lines lead to 0.5 point > > > losses. > > > > > > i wonder if this kind of greediness might, however, be useful for > > > selecting, > > > say, the first move or two in a 9x9 game. the thinking here is that > > > since the > > > endgame is essentially noise at this point, you might as well be greedy > > > before tactics become an issue. that's probably faulty intuition, though. > > > > > > on another note, has anyone just let their MC code rip for a day or two > > > (or > > > maybe a week or more) on the first move alone? i would think that if you > > > ordered the distribution of the resulting list, it would give very good > > > information > > > about how well MC acts as a board-eval function. (i.e. turn off all book > > > lines, > > > turn off all rules about not playing on the first line or two early in > > > the game, etc. > > > etc. turn off all heuristics related to the opening and then print the > > > distribution > > > over the board). what are the top, say, 10 moves on a 9x9 board and how > > > are > > > they distributed, and the top, say, 40 moves on a 19x19 board along with > > > their > > > distribution? if you fold board symmetries into your search, i suppose > > > that you > > > can get a factor of 8 here. > > > > > > my thinking is that if it's anything other than a very smooth > > > distribution among > > > the top moves, that's a good indicator. > > > > > > s. > > > > > > > > > > > > > > > > > > > > > ____________________________________________________________________________________ > > > It's here! Your new message! > > > Get new email alerts with the free Yahoo! Toolbar. > > > http://tools.search.yahoo.com/toolbar/features/mail/ > > > _______________________________________________ > > > 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/ _______________________________________________ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
