On Thu, Jun 16, 2011 at 8:11 PM, Hideki Kato <[email protected]> wrote:
> > Don Dailey: <[email protected]>: > >On Thu, Jun 16, 2011 at 2:58 PM, Matthew Woodcraft > ><[email protected]>wrote: > > > >> Don Dailey wrote: > >> > The concept I'm trying to refute isn't whether some particular program > is > >> > not scaling but the concept that "MCTS don't scale", as if the whole > >> concept > >> > is seriously broken. > >> > >> I think you would have done better to try to refute that concept by > >> following up to a message in which someone was expressing it, instead of > >> following up to Hideki's message in which he said nothing of the sort. > >> > > > >Here is what he said: > > > >That's simple (at least I think). The scaling law converges at some > >point where the speed (or more thinking time) benefits little. Zen19D > >already reached that point and Zen19S's rank is lower than Zen19D's. > > > > > >Maybe I misunderstood, but he talked about the "scaling law converging at > >some point" which sounds like a "law" (that is applied uniformly) and > then, > > "Zen19D already reached that point" which sounds to me like he was > >complimenting Zen on being so strong it has reached the limits of the "law > >of scalability" already. > > > >If I misunderstood it's because he didn't express himself very clearly. > > Ok, I've omitted why I wonder the scalability of MCTS in my post to > avoid multiple issues in one reply. Below is the reason. > > UCT's global convergence theorem was proved with the assumptions, the > stocastic process is stationally and i.i.d (independent, identical > distribution). These could be not held in the game of Go. > > Also, the proof requires the factor of the exploration term has to be > greater than some constant which we have no way to know in practice. > > So, my conclusion is that we have no way to develop guaranteed > correct (in Don's sense) MCTS programs now. > > #Sorry for unclear posts. Japanese way of thinking is not so > logical nor clear as Westerns way, probablly because my mother language > is basically not logical but emotional and my English skill is not > enogh. I'm doing my best but still seems not clear enough. Please > point out if you find somthing not clear. > Fair enough. I'm pretty sure the algorithm or some modification of it is scalable, and it may be that as hardware continues to advance (at least for as long as that continues) constant adjustments might be needed to keep things moving. Things that could in practice be a part of the program now. So even if it's not technically scalable due to some minor implementation details I think the general concept is. Maybe another scalability study is in order? I'm not volunteering, but it might be good to see real data instead of anecdotes and speculation on either side. With multi-core so common it might not be difficult getting a serious amount of data with a little cooperation. Don > > Hideki > > > > > > >> > >> I wouldn't normally bother mentioning it, but I think your message was > >> rather rude (particularly the "I'm trying to save you from looking > >> stupid in 10 years" part). > >> > > > >Ok, this was a rude on my part and I apologize. I could have been a bit > >more delicate for sure. > > > >Sometimes you have to tell someone they have body odor as a kindness, even > >though you KNOW it's not going to go over too well. This was given in > that > >spirit - I AM trying to get people to stop saying these things that make > >them look naive now so that people won't reflect on this in a few years > and > >laugh at. I'll bet Bill Gates wished he been warned when he said, > "640K > >ought to be enough for anybody" although he denies saying that. Would > it > >have been cruel, or kind to tell Gates that he is going to sound stupid > >later if he says that now? > > > >Don > > > > > > > > > > > >> > >> -M- > >> _______________________________________________ > >> Computer-go mailing list > >> [email protected] > >> http://dvandva.org/cgi-bin/mailman/listinfo/computer-go > >> > >---- inline file > >_______________________________________________ > >Computer-go mailing list > >[email protected] > >http://dvandva.org/cgi-bin/mailman/listinfo/computer-go > -- > Hideki Kato <mailto:[email protected]> > _______________________________________________ > 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
