On Sun, 2008-08-10 at 19:32 -0400, Robert Waite wrote: > Well... I think I have hunches just as you do. And I think we both > express our hunches on here. > > Diminishing returns is not really my theory.. I am just looking at > alternative ways of viewing the datapoints. Let's say you have two > computers and both of them focus only on solving local situations. At > first they both play around the same level. Then you scale one of them > in some way and mark the trend. We now know that one of them scales a > certain way when solving local solutions. If you then take that > computer and put it against a person.. the person is no longer just > thinking about the local solution. They are thinking about strategy, > they might make certain moves that test the computer's goals.. you > have a whole different situation. > > The little green men reference looks like a dangerous use of Occam's > Razor. In this case... someone says there are little green men under > his house and shows me a bunch of datapoints to suggest it. I don't > have any datapoints against it so my point is automatically > invalidated? It seems there is a little more detail to Occam's Razor.
The little green men was more about burden of proof, not Occam's Razor. As far as Occam's Razor is concerned it does not deal with proof or disproof. It's a general principle and sometimes it doesn't work because things sometimes really do have complicated explanations. It's another way to say, if it looks like a duck, walks like a duck, quacks like a duck, don't keep insisting that it's a dog. Have you seen the contortions that the flat-earthers go to in order to explain away the facts? They continue to produce a stream of datapoints to suggest they are right and that doesn't prove they are wrong, but I'm not going to waste my time trying to refute them. > > When I saw "proven to be scalable", my first thought was that it was > proven to be somehow practically scalable in order to beat a human or > perhaps to solve the game. You even mentioned that God would draw with > the computer. That kind of scalability seems related to solving the > game, not to beating a human. I saw no evidence that it has a > practical scalability to solve the game. And this seems like the kind > of problem that could be intractable. Practicality is an issue. Are you seriously trying to make an issue out of whether a mini-max search or MCTS can "solve" 19x19 go in any practical sense? If that's what you've been talking about, then I think we both wasted our time. Let me set your mind at ease, I agree with you. I don't know what that has to do with the price of tea in China and I didn't know you were hung up on that point. Anyway, that has nothing to do with anything I was talking about. I thought we were talking about whether we had reasonable hope that a go program could be equal to the best players withing a very few decades. > Now the topic has moved to scalable to beat a human and I disagree > with the interpretation of the data. We are both interpreting data. > Your data doesn't count as a theory.. where you reduced my theory to > one that has no data. We are both interpreting the same data. What are you talking about? The only thing I see going on here is that you are playing devils advocate and throwing out stuff without much basis. We are both doing some conjecture and that's ok with me but you seem to be going out on a much farther limb than I am. You are the guy that says the duck is a dog. I agree that there is a chance you could be right, there is a slight chance that it could be a highly deformed dog with a vocal abnormality. > Diminishing returns was just an example of something that could be a > roadblock. I was questioning how this necessarily scales to humans. It > seems more data is needed from MC-programs vs. humans to make a > rigorous theory of scalability. So far.. the only scalability that > seems proven is a case for solving the game... not beating humans. You cannot solve the game without beating humans along the way. > There is some point between that would most likely in my opinion lead > to humans being beaten.. some amount of calculation before you solved > it.. but the shape of this curve is something I am unsure of. It's shaped like a duck. It starts out with a gentle slope, then it violently turns upward (forming the head) where it flattens out near the top, then begin the gradual downward slope to the beak. Then it doubles back on itself to form the lower beak before it begins forming the breast, moving forward once again. That's my theory anyway. Can you prove me wrong? Occam is wrong about this one! > It doesn't seem that unreasonable to question if there is a practical > scalability. It does seem unreasonable to me. But it's certainly ok to have some questions. I'm pretty confident that it will work the same as in every other game - extra CPU power makes the programs stronger against ALL opponents, not just humans. I think the word you are looking for is transitivity. You believe (or at least theorize) that there is little transitivity between humans and computers. In other words you believe that a computer program might improve many stones against other computers (for instance) but none or very little against humans. This has been talked about many times on this group. My personal view is that intransitives exist, but can only be pushed so far. If you can get 10 stones improvement in a GO program against other go programs, you are still going to get a LOT of stones against humans, although it may not be exactly 10 stones. You can get a pretty decent improvement if you tune against a single opponent without it translating to much against other opponents. But not several hundred ELO. It's like a rubber band that cannot be stretched but so far, it always wants to come back. Take an 1800 ELO rated chess player for instance. It's possible there are 1900 or 2000 rated players who would lose to him in a long match because of playing styles that match up well for the weaker player. But it's far less likely that a 2400 player would be troubled by this, even if his style was not ideal for beating the 1800 player. - Don > > _______________________________________________ > 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/
