On Sun, 2008-08-10 at 15:19 -0400, Robert Waite wrote: > > Hmm.. I dunno.. I think there are a lot of ideas floating around but > some miscommunications. > > So the aim is to devise a computer that will beat the strongest human > players of go. > > I hear that "Monte-Carlo with UCT is proven to be scalable to perfect > play". It seems that this is essentially saying... that as the sample > size for this technique grows to infinity.. you will approach the > accuracy an algorithm that has solved go (in the sense that 5x5 was > solved)... kind of like creating the entire game tree. That this curve > approaches perfect play as you increase the samples to infinity. Same > goes for drawing out the entire game tree. It just seems that MCwUCT > is a lot easier. > > This however speaks nothing about the rate at which it approaches > perfect play as you increase the sample size. I didn't see anything in > the papers I have read about this. Which brings us to what our aim > is.. and that is to beat human players at go. Nothing has been proven > yet about practical scalability... which is what we would like.
I don't know how you can say that. The empirical evidence is overwhelming that this is scalable in a practical way but more importantly it's been PROVEN to be scalable. If you throw the word "practical" in there then you are no longer talking the language of mathematics, theory and proofs so please don't ask for a "proof" of "practical" scalability, it makes no sense. > Scalable in the sense of approaching infinity alone does not prove > that it is not intractable. MC will never "solve" the big board in a "practical sense." We are of course talking about the issue of scalability in a practical game improving sense. You can't have it both ways - the tool we use to predict performance is an understanding of it's scalability properties. That has been proven but you want to say it's not relevant and that we have no way to estimate it's chances against humans. But we DO have a way to get a rough estimate. I think many just don't want to believe that it's possible to beat a human - their brains cannot get used to the idea that it will probably happen within their lifetime. > And it was said that the Mogo devs said that a double-strength version > beats the other one with 63%. As they mentioned... ideally... this > would mean about 30 years. But there could be a point of diminishing > returns as it relates to beating a human. There clearly is diminishing returns, even at very weak levels but you probably cannot measure it. I think it's very likely that the diminishing returns curve will be very very gradual for a long time to come, well beyond the point of achieving the top human levels. I believe for many this is a matter of credulity. It was like this in chess, we could not accept the possibility that computers could really get that good. And in GO we are so far away still that our brains have to make up reasons why it can't happen. When our realities don't match our belief systems, we balk. But you can look at it another way - humans are not that good at the game. Compared to you and I, they may seem like god's, but they are fallible and weak in the perfect game sense. Grandmasters in chess, in many ways fell from grace when it became necessary to play odds matches with computers in order for them to have a chance. When I was a boy reading chess books I practically worshiped these immortal masters of chess. The computers now expose their errors all the time, they are just humans and are no different from you and I, they just play a few hundred ELO stronger. If you take them off the pedestal, you can think more rationally about it. - Don > > When you say that is it proven scalable to perfect play... it is like > saying that we know that if you create every possible game... and have > a database that can access it well... you can get to perfect play. > This does not help us actually do it. > > _______________________________________________ > 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/
