There is no "wall", you are just seeing things from too local a perspective.
When you look at a small gradual slope from a distance, it looks like a huge mountain and it's easy to overlook the fact that there is a pathway to the top. That is what we have here, a very long way to the top. We do not have a wall, we have a mountain. Darren mentioned that I would say the algorithm is scalable. Apparently this is a reference to Moores law. Unfortunately, Moores law may not hold, and based on what we currently know about physics, we cannot expect it to continue forever. I think what happened in computer chess, is that the problems we faced could legitimately be viewed as "walls" and were proclaimed to be so. In magazines like personal computing there were experts who were saying that no computer would ever reach expert level, which is a measly 2000 ELO. At the time, it could be said with reasonable confidence that this was true. It was truly a wall and it was almost a self-evident truth, considering the computing power available and we ever hoped to have available. But it wasn't really a wall, it was a long climb up a mountain. I think many of you on this forum do not have the advantage of historical perspective here, but I was there the whole time and it literally took decades for computers to get where they are now, and it was VERY gradual. So we had a mountain, not a wall, and mountains can be climbed. None of the imagined congintive barriers that would prevent a program from ever playing 2000 level chess ever existed. I'm being serious here, it was claimed even that category 1 (or class A play) was beyond computers no matter how fast they got. And now of course we recognize that as complete foolishness, but at the time it sounded pretty reasonable. If Moores law WERE to continue to hold, which none of us expect, then a computer go program would be world champion in 200 years, probably much sooner. In 100 years I estimate 66 doublings. 66 doubling is 73,786,976,294,838,206,464 times faster speed. Do you honestly think that would not make a go programs much stronger? Now if you want to argue that Moores law won't hold, then we have no argument. I agree with you completely. But if you are arguing that there is a conceptual barrier that makes any increase in computing power worthless, I have to say you are just being shortsighted and you only see the mountain in the distance and even then you see it 2 dimensons, as this thing sticking straight up in the air (a wall) instead of a slope to a peak. But here is someting interesting: In the case of computer chess it has been estimated that the progress in software has been roughly the same as the progress in hardware. Modern chess programs are truly amazing, and not just a result of faster hardware. There is no reason to think that this won't be true of computer go. (In fact, I believe go software has even more potential than in chess due to the fact that problem is conceptually more interesting.) So if we were to equate this to hardware, as unlikely as it may seem, we might expect the equivalent software improvement of 66 doublings in 100 years. Of course since 100 years is too far away, we tend to see this as a wall because as impatient people we say, "I want it yesterday." So if you want it yesterday, then for all practical purposes it is a wall. I have one more comment below on what David Fotland said. On Wed, Jun 10, 2009 at 5:10 AM, Darren Cook <dar...@dcook.org> wrote: > Olivier Teytaud wrote: > > As programs improve, my humble opinion is that computational power > > becomes less and less efficient as now the main trouble is the > > systematic bias for some positions (semeais or too complicated > > life-and-death situations) - parallelization is not required. > > I wondered about this when checking my facts for the Shodan Go Bet page > (http://dcook.org/gobet/), as it seemed CrazyStone, ManyFaces and Zen > are getting similar results on, say, 8-cores as Mogo was getting on big > supercomputers. (With the usual disclaimer that sample size is small so > the error bars are big...) > > And David Fotland commented [1] with effectively the same thing: current > MCTS programs won't be European 1-dan no matter how much computer power > they have. (Instead algorithm issues need to be fixed.) David Fotland may have said that, but he didn't mean that. If I gave David Fotland a computer that was 73,786,976,294,838,206,464 times faster than what he has now, do you honestly believe he would not be able to produce a European 1-dan strength player? If he cannot do it, then I can. Surely, what he really meant was that he believes no computer in the forseeable future would be enough. I would like to remind everyone here that our track record as humans of seeing very far into the future is pretty lousy. - Don > > > Don will jump in and say but the algorithms scale, this cannot be true. > But, while that may be the case, perhaps we can say that they are > hitting a wall in their observable playing strength against non-MCTS > players (such as humans) at higher levels. In [2] I touched upon how the > nature of the game changes at higher levels, and how scaling results > obtained between weaker players may not apply at those higher levels. I > was talking about pure random playouts in that article, but the > systematic bias Olivier mentions can lead to the same problems as no > bias at all... > > I'd love to hear what David and others think about the kind of > algorithms (or knowledge?) that are needed to get past this wall. > > Darren > > > [1]: > http://answers.polldaddy.com/poll/1614035/?view=results > > [2]: > http://dcook.org/compgo/article_the_problem_with_random_playouts.html > > > -- > Darren Cook, Software Researcher/Developer > http://dcook.org/gobet/ (Shodan Go Bet - who will win?) > http://dcook.org/mlsn/ (Multilingual open source semantic network) > http://dcook.org/work/ (About me and my work) > http://dcook.org/blogs.html (My blogs and articles) > _______________________________________________ > 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/
