fminsearch2

Modified Unconstrained Optimization with better starting point selection

Description

This is a (small) change to the current fminsearch routine to improve convergence.

The current fminsearch selects n additional simplex points based on a small change in each variable. This implementation uses an adaptive procedure to select these n points while maintaining the correct relative distance to the centroid of the remaining points.

Changing these initial points gave 11% fewer runs and reached the correct solution 6% more on the test suite by More et al.

• Required Products: Optimization Toolbox
• MATLAB release: MATLAB 7.4 (R2007a)
• Other requirements: Should be no different than the original fminsearch

[John D'Errico]

This is a copy of fminsearch, but the help has not been fixed to show that you should use fminsearch2. (I don't know if there are any legal issues with submitting MathWorks code on the file exchange.)

I did not get that deeply into the code, merely tried a couple of various scalings of the Rosenbrock function. For most of my tests, fminsearch2 took exactly the same amount of function evals and iterations as did fminsearch. Some additional testing did find one case where fminsearch2 was 30% more efficient. I stopped at that point.

[Duane Hanselman]

If this is really better than fminsearch that ships with basic MATLAB, then proposed it as an improvement to TMW. TMW might not be in favor of posting their code having only a few bytes of different code. Please remove this submission.