ivan dubois wrote:
> Thanks !
> I think I agree with everything you said.
> After some reasoning, I changed my opinion about the practical scalability of
> Mogo. It does have some missconceptions about eyes, but as you said, it can
> be corrected by the UCT part. I tried to find a situation where the UCT part
> does not help until the situation comes to a conclusion, with no success. So
> you are probably right, maybe I will give some more thought about it, or not
> :)
>
> But I stay on my position that the theoritical "infinite scalability of UCT"
> doesnt give any hint about whether or not it is scalable in practice. I think
> you made my point clear : From a pure theoritical point of view, it has no
> advantage over a simple mini-max solver. Given that a mini-max solver is NOT
> considered scalable in practice, I see no reason why people would see this
> property of UCT as an indication of its scalability.
>
This is still nonsense. UCT in actual real world "PRACTICE" responds
dramatically to more hardware, how can you say it's not clear whether
it's scalable in practice?What is practice to you?
As far as mini-max, that is also scalable in the real world that I live
in. Aya is an example of a reasonably good mini-max player that
scales.The theoretical properties of scalability hold in actual
practice.
If you mean that we will never see in our lifetimes real world hardware
fast enough to see it play perfectly, everyone already knows this, it's
not any kind of revelation. (However I'm surprised at how many people
get hung up on this point - thinking it is somehow relevant as a proof
that the idea of search is totally useless.)
The property we are interested in is that we can always expect an
increase of playing strength by doing one or more of the following things:
1. let it think longer.
2. let it think faster in the same amount of time.
3. continue to improve the knowledge.
4. continue to improve the efficiency of the whole system
"infinitely scalable" is a poor term since there is a finite limit.
The finite limit is perfect play and as I said, we don't actually expect
to achieve that.
- Don
> Ivan
>
>
> - Message d'origine
> De : Weston Markham <[EMAIL PROTECTED]>
> À : computer-go
> Envoyé le : Mercredi, 23 Janvier 2008, 0h41mn 08s
> Objet : Re: Re : [computer-go] Is MC-UCT really scalable ... is a troll
>
> (I typed the following up earlier today, before other people cast some
> doubts about what "infinite scalability" means. Perhaps it is
> helpful.)
>
> I think that Alain was specifically referring to a property of UCT,
> whereby given that a winning line of play exists, the probability that
> the algorithm has discovered one such line & settled upon it
> approaches 1 as time approaches infinity. I believe that this has
> been proven. And no, this theoretical result does not represent any
> intrinsic advantage over a calculation of the minimax value. (Someone
> should feel free to correct me if there is a stronger theoretical
> result that I am not aware of, though.)
>
> I think it is important to understand that unless you restrict UCT
> from exploring some correct lines of play, this "infinite scalability"
> will still be true. And, it is true regardless of what moves have
> been preferred/pruned/whatever within the MC playouts. (Or whatever
> one does at the leaf nodes; it need not be MC at all, as long as the
> UCT tree eventually gets expanded as the node is revisited.)
>
> This is not to say that this particular property of UCT is sufficient
> in order for us to claim that UCT is scalable in practice. My point
> is rather that noone is claiming "infinite scalability of MC", but
> rather scalability of UCT. When you combine that with the fact that
> UCT has been very successful at 9x9 go, (and scalable across a wide
> range of time limits) it seems to be reasonable to expect more of the
> same in 19x19. (Which is what concerned the original poster.) Also,
> one should expect that the UCT portion of MC-UCT will tend to
> eventually "fix up" any systematic errors that are made by the MC
> playouts.
>
> I have one other point I'd like to make, in regard to "light" versus
> "heavy" playouts: Even "light" playouts do not actually use a uniform
> distribution, due to the quasi-eye rule that is generally used. I
> think that there is every reason to expect that other "nonuniform"
> playout policies further outperform "light" playouts in every
> practical way. Granted, it is possible to introduce "severe
> misconceptions" when one incorporates knowledge into one's playouts.
> But even in that case, one can still fall back on the fact that UCT is
> cleaning up after one's mistakes: the eventual behavior, given enough
> time, is still perfect play. (And of course it is not as if people
> blindly adjust the monte carlo policies without checking the revised
> versions for "severe misconceptions".)
>
> Weston
>
> On
