On Thu, Jun 2, 2011 at 9:00 AM, Jacques Basaldúa <[email protected]> wrote:
> Don Dailey wrote: > > > I don't understand your concept of "wrong direction." When > > you expand the tree for ANY move you get more information about > > how good or how bad that move is, so it's always helpful (on > > average) to have more time. > > I think Hideki's argument is about: more simulations won't > reach the right move in real-world circumstances. > > If there is a skillful move that requires a precise sequence and > the playouts get it wrong, the tree will not explore that direction > frequently enough to find the skillful sequence by itself. It may > take a million years and more RAM that is available on the planet. > But that has nothing to do with anything relevant here. The fact of the matter is that even the very best MC programs suffer from positions where a million years of computing will still give the wrong results. This is even true in computer chess. If you believe that we have some method available in go that always finds the right move in a modest amount of time, then we have almost solved the game. But we haven't and the fact remains that if you do more playouts your program will play stronger. And if you have 10x more time it will be non-trivial, even if your playouts suck. The only wall is perfect play. You cannot get past that one and any sensible scalable program will asymptotically approach it with power. However, I think it's clearly true that the quality of the playouts has a profound impact on the strength of the program and even the rate of improvement as you think longer. I think it likely that with uniform random playouts a program might achieve 1 Dan but the same program with superb heavy playouts might achieve many Dan. > > That is a fact and it can be shown by examples, but that does not > contradict your point. These positions exist, but there will also > always be positions that fewer simulations get wrong and more > simulations get right. So scaling will always maintain, unless > there are bugs or RAM limits. When the tree stops growing, its > leaves are basically tree-less MC evaluations and it is known that > they do have an asymptotic behavior. > Agreed. Of course you can find examples that prove that heavy playouts can solve positions that light playouts have no chance of solving in reasonable time - but I'm not arguing that point. Don > > > Jacques. > > > _______________________________________________ > Computer-go mailing list > [email protected] > http://dvandva.org/cgi-bin/mailman/listinfo/computer-go >
_______________________________________________ Computer-go mailing list [email protected] http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
