[computer-go] .. if Monte-Carlo programs would play infinite strong

[Don Dailey]

Yes, I agree with the point you are making.  Random play is a relatively
good evaluator, but it is not a great evaluator.   And it's very weak at
tactics.   Letting it do a lot of simulations does not cause it converge
to the correct value.

But the current breed of MC computer players do not have a fixed tree -
they continuously expand the tree in best first manner.  

Perhaps we use the wrong terminology when we call these MC players
because they are hybrids.   But I think that's understand now when we
say "Monte Carlo player."

- Don


On Sat, 2006-11-25 at 10:05 +0000, Jacques Basaldúa wrote:
> Maybe I did no explain my point well enough.
> 
> The problem with infinite is that we get a better approximation to a 
> wrong value.
> 
> With few simulations we get that value with, say 1/10 error. With an 
> astronomical amount
> of simulations we get the same value with an error of 1e-200, but it's 
> still wrong!. It is
> proved that simulating a go position converges, but it does not converge 
> to the same
> value as if the position was played by perfect players, it only 
> converges to the asymptotic
> limit of random play.
> 
> I am not an MC developer, but as far as I know, UCT keeps a limited 
> (i.e. n-ply) tree
> in memory and intentionally unbiasses the nodes to make the convergence 
> faster, that
> does not change anything, assuming constant tree size.
> 
> A simple test :
> 1: after 100 simulations, choose the highest number in (0.96, 2.1, 3.18)
> 2: after 1e9 simulations, choose it in (0.9999999, 2.0000001, 3.000001)
> You chose the same value (= same move).
> 
> That's why, I insist, if you don't increase the size of the tree and 
> only get a better
> approximation to a wishful but frequently misconceived value (the limit 
> of random
> play) witch is *not* a good evaluation of the game, you don't 
> significantly improve
> your play. Of course, if you increase the tree, you reach perfect play, 
> that's not
> the point.
> 
> Jacques.
> _______________________________________________
> computer-go mailing list
> computer-go at computer-go.org
> http://www.computer-go.org/mailman/listinfo/computer-go/