On Fri, 2007-04-13 at 19:52 +0900, Darren Cook wrote: > > http://greencheeks.homelinux.org:8015/~drd/public/study.jpg > > ... > > I'm actually testing 2 programs - both of them UCT style go > > programs, but one of those programs does uniformly random > > play-outs and the other much stronger one is similar to > > Mogo, as documented in one of their papers. > > Hi Don, > Can you describe the implementation of heavy and lite in more detail; > especially lite? > > I've been trying the libego program "out of the box", and am up to > 200,000 UCT playouts, but still gnugo 3.6 on level 6 is winning 10 out > of 10. According to your chart, lite with 128K playouts should be about > 50 ELO points higher than gnugo 3.7 (on level 10 I assume?). (256K > playouts should be 130 points higher.) > > I'm wondering if there is a standard set of UCT enhancements that > everyone is doing? > > TIA, > > Darren
If 200,000 play-outs is being beat, something is broken. Even a bad implementation should do better than that. Does libego provide a full UCT implementation out of the box? The lite version I am testing has one enhacement that makes it not uniformly random. There is code to make it prefer captures in a probabilistic way. I think I just test N moves, and if one of them is a capture I play it, otherwise I choose a move randomly which doesn't fill an eye. But this enhancement isn't needed to get it up to the level of gnugo 3.6 There are issues with things like suicide, eye-filling and so on that can effect the strength - different people claim different things about this. I never allow suicide, I test only for simple ko in the play-outs. I don't know how much impact variations of these things have. One thing that I think most people are doing is related to how you select a move once you have finished the search. Do you pick the highest scoring move? Sometimes the highest scoring move has very few simulations in that part of the tree and the simplest and safest thing to do is pick the node that has the highest number of simulation under it. This move is usually the highest scoring move, or has been the highest scoring and probably will be again. Another thing you can do is to delay the search to see if the new move sticks. That's the only thing I would consider an enhancement - I don't know how much strength it contributes but it does help. - Don > _______________________________________________ > computer-go mailing list > [EMAIL PROTECTED] > http://www.computer-go.org/mailman/listinfo/computer-go/ _______________________________________________ computer-go mailing list [EMAIL PROTECTED] http://www.computer-go.org/mailman/listinfo/computer-go/
