It uses a variant of Powell's method, which involves sequentially
optimising one parameter at a time, but with slight modifications to
make the search more efficient.
Each one parameter search involves bracketing a minimum (ie finding
three points where the solution for the middle point is better than that
for the two either side.
Then it will fit a quadratic through the three points, and try a new
point that corresponds to the minimum of the quadratic. If this is
better than either of the current three points, then the algorithm will
take the three best points and continue homing in on the solution.
In situations in which the quadratic approach does not provide a good
estimate for a new point, the algorithm will try a golden section search
strategy instead.
All this stuff is described in Numerical Recipes, and is one of the
simplest optimisation strategies available - as it does not require
derivatives.
Best regards,
-John
On Mon, 2009-08-03 at 22:37 +0100, Henrique Amaral wrote:
> Hi SPMers
>
> The coregiter algorithm use the descendent gradient in SPM2? What is the
> optimisation used???
> Thanks
>
> Henrique Amaral
> USP - Ribeirão Preto
>
--
John Ashburner <[log in to unmask]>
|