> So I wouldn't be surprised at all if at some point you'll see a > marriage of the best ideas of traditional Go programs and Monte- > Carlo / UCT. In fact, this is most likely already happening as these > Monte-Carlo programs use algorithms / ideas from the traditional > programs for tactics, pattern-matching and possibly others.. And I go > out on a limb to predict that at some point "heavy playouts" will use > information like territory, eye-space and group-strength to guide > their playouts just like the "traditional" programs do. And that > using fixed 3x3 patterns will be a passing fad.
You don't even have to go out on a limb ;-) One detriment to this is still that increasing the strength of playouts does not make them necessarily better. This might cause a schism in the kind of information that is needed between classical approaches and Monte Carlo. So I do not think both approaches will totally converge. > This comes to my first point. Optimizing early in a project is like > listening to the devil. It eats up a lot of time, the visible > progress is gratifying but in the grand scheme of things it's not all > that important to do early. I implemented it using pseudo-liberties > because... uh, well, because that's what everyone seemed to be doing > to get high numbers of playouts per second. But I already start to > doubt using pseudo-liberties are all that useful. Is anybody still > using pure-random playouts for anything? As soon as you start to do > anything more, pseudo-liberties are pretty useless. I agree here, and I made the same mistake. > But the speed of the random playout becoms less and less important > with heavy playouts. This I don't understand at all. The improvement curve for being faster isn't different with heavy than with light playouts. > Next I was thinking about another subject that got some attention on > this list, the mercy rule. It seems to save about 10% in the number > of moves per game (on 19x19) and result in about 20% gain in > performance. This discrepancy is most likely due to the fact that the > administration, whether using pseudo-liberties or real, is much > slower towards the end of the game because you have more moves that > merge chains. And those 10% moves it saves are of course always at > the end. So is it relevant? I don't know whether heavy playouts will > be slower towards the end of the game or not. Possibly yes, as more > moves made will have a small number of liberties that will need > tactical analysis. I'd say that generally reducing the move-count is > a good thing whichever method one uses. Possibly at a later stage > more sophisticated methods can be developed to abort a game early. Possibly the mercy rule is a premature optimization just like psuedo- liberties :-) > Next I allowed suicide only for White. You allowed single-stone suicide too, I guess? This is obviously bad since it will happen constinously near the end of an MC playout. It will be like one side is forced to pass 50% of the time and the other side not. -- GCP _______________________________________________ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
