On Thu, Jun 2, 2011 at 3:48 PM, Jacques Basaldúa <[email protected]> wrote:
> Don Dailey wrote: > > > Are you trying to say that heavy playouts are better? > > Who is going to argue with that? I agree completely. > > Are you trying to make the point that there are very simple > > to understand positions that computers cannot easily solve? > > I agree with that. Are you trying to say that heavy playouts > > can solve many types of common positions orders or magnitude > > faster than light playouts? I agree with that. > > Are you trying to say uniformly random playouts suck? > > I agree with that. > > I do not pretend to argue. Just to clarify ideas and read what > others have to say. And of course I agree on all that. > I'm not really directing this to any specific individual, sorry it came across that way. > In self play all MCTS programs scale. Everybody agrees and it > has been tested empirically. Intuitively: If we admit that 2000 > sims is better than 1000, since nodes in the tree are trees > themselves, it is clear that no matter how many million > simulations we play, there will always be nodes with 1000 visits > and they would be better evaluated if they had 2000. The entire > tree relies on the correct evaluation at the nodes so the entire > tree benefits of more sims. > > A different question is: Can a really weak program, say vanilla > MCTS with uniform random playouts, just no eye filling (no RAVE, > no progressive widening) reach the strength of, say Aya, with 2500 > sims (KGS 4 kyu) in 19x19 ? > That's not an interesting question. You can just run a program like this and get your answer as you have mostly specified the level and the algorithm. I don't know the answer, but 4 kyu seems pretty strong to me for a program that only uses uniform random playouts and no tricks. > The answer is: > > Theoretically: Yes. > In practice: No. Not with a trillion sims per move. > Uniformly random sims will often send the program down an incorrect pathway, but the program eventually discovers it's error (assuming there is SOME exploration) and will find better moves. A trillion sims is a LOT of sims and I believe it is enough to get above 4 kyu, which is not a very high level. But I don't really have a good way of estimating this level. I do believe that with the quality of the sims there has to be some adjustment to the tree search algorithms. If the sims are horrible you cannot depend on them as much to direct the shape of the tree and visa versa. So I would say that if such a program is allowed to be properly tuned and gets to do a whopping trillion playouts per move, it's going to be stronger than 4 kyu. I believe that even an alpha/beta search (perhaps like Aya was) is doing a trillion nodes per move, it is going to play a good game. To be sure, you would still be able to find some simple position that it screws up big time. But isn't that also true of the 3 and 4 Dan programs we now have? So why can't we have a 4 kyu program that plays stupid moves too? > > You probably don't disagree since that is implicit in "heavy > playouts can solve many types of common positions orders > or magnitude faster than light playouts". > Heavy playouts is like turbo charging the program. The difference between a uniform simulation and heavy playouts is not just enormous, but grows with each doubling. In practical terms, you cannot have a strong program without heavy playouts and in 10 years the majority of progress will be with the playouts, not the search. The quality of the software (if it's like in chess) will grow at least as fast and probably faster in GO as the hardware. Of course if we are seeing the limits of Moores law, that is even more true, but I don't think we are (they said that 10 years ago, didn't they?) > > Note that this question is equivalent to: Would the current > version of Zen become a pro just with hardware evolution? > When you ask questions like this you open things up to confusion and argument because you didn't ask a well formed question. You did not specify a time period or any kind of rate of hardware evolution. Are you just asking for a prognostication? With infinite hardware evolution (assuming it will always reach some arbitrary speed if you wait enough years or centuries) then the answer is yes. But at some point the limits of physics must rear it's ugly head and the hardware progress with be much slower. So it's not possible to give any kind of estimate that is attached to reality. I can only guess about this, but I think we are going to be surprised by how much a 100x faster computer gives us. If you had asked this exact question 35 years ago with respect to computer chess, most reasonable people would have said that having a computer 10,000 times faster would probably only give you a hundred to two additional ELO. I'm not joking, this is how it was. This happens because when we see a really big problem, we imagine that it's even bigger than what it is. It was very common back then to show people problems from real games that no computer imaginable would ever be able to solve. And yet now they are solved trivially. I don't mean to be disparaging, but it shows how stupid and naive and shortsighted we all tend to be at times. The Zen question won't happen because the software (in my opinion) will evolve much faster than the hardware. But they will both be very important. We will get a pro playing Go program decades sooner if we don't wait for hardware to run Zen on, but instead we make constant advances. 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
