[computer-go] How to improve my minimax speed?

[Lars Nilsson]

On 11/14/06, Eduardo Sabbatella <eduardo_sabbatella at yahoo.com.ar> wrote:
> Thats impossible to said without error %.
> That error could lead the alfa-beta algorithm to prune
> an important move on a 'bad' board configuration for
> our evaluation function.

See attached image. Alpha-beta would never examine the node marked X,
nor any below it, given the calculated value 2 of its sibling. You
only need to know that you can at least get a score of 8 by picking
the top branch, while your opponent can guarantee at least a maximum
of 2. So you pick the top branch, without considering what X might be,
because your opponent won't let you go down that road unless it's even
worse than 2.

Now just imagine the alpha-beta algorithm working on multiple levels,
throughout the tree, and you'll find yourself with a whole lot less
nodes to evaluate (given a reasonable move ordering).

> Its impossible to decide if a move worths NN points in
> GO withour error.
> (that would be the "ultimate" engine by itself).

Please keep in mind that you get exactly the same error for each
evaluation when using minimax. Regardless of the quality of your
evaluator, if you use the same one for both alpha-beta and mini-max,
you'll get the same best move.

Lars Nilsson
-------------- next part --------------
A non-text attachment was scrubbed...
Name: alpha-beta.gif
Type: image/gif
Size: 5390 bytes
Desc: not available
Url : http://computer-go.org/pipermail/computer-go/attachments/20061114/d7707e45/alpha-beta.gif