[computer-go] analysis of UCT and BAST

[Jason House]

In looking at the paper again, I cheated and looked at the graphics and the
conclusion.  For those of us who are concerned with comparison to UCT, I
believe that it's off the chart to the left, but would likely match the
plateau.  Initially, I was looking at the plot as if it was time series
data, but it really isn't.  I think the Rn through me off since I didn't
read all the way to n=10^7.

The conclusion says "it seems important to use true upper confidence bounds,
in order to avoid bad cases as illustrated in regular UCT."  That caught my
attention, but I didn't put in enough effort this time around to figure out
what it was talking about.

On 5/30/07, Remi Munos <remi.munos at inria.fr> wrote:
>
> I have updated the BAST paper, providing additional comparison with UCT,
> as
> suggested by one person in the list. See:
> https://hal.inria.fr/inria-00150207
>
> Basically, in the considered examples, BAST with an appropriate smoothness
> sequence performs always better than both UCT and Flat-UCB.
> Now, the best smoothness coefficient is always smaller than the constant
> corresponding to the true smoothnes of the tree, and sometimes very close
> to
> 0 (which corresponds to UCT).
> Since go is way more complex than the problem of optimizing a simple 1-d
> function, I would say that no theoretical work could predict whether BAST
> would do better than UCT or not, in Monte-Carlo-go.
>
> Best, Remi
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://computer-go.org/pipermail/computer-go/attachments/20070530/28d68f9a/attachment.htm