Enhancement of the downhill simplex method of optimization

The downhill simplex method of optimization is a "geometric" method to achieve function minimization. The standard algorithm uses arbitrary values for the deterministic factors that describe the "movement" of the simplex in the merit space. While it is a robust method of optimization, it is relatively slow to converge to local minima. However, its stability and the lack of use of derivates make it useful for optical design optimization, especially for the field of illumination. This paper describes preliminary efforts of optimizing the performance of the simplex optimizer. This enhancement is accomplished by optimizing the various control factors: alpha (reflection), beta (contraction), and gamma (expansion). This effort is accomplished by investigating the "end game" of optimal design, i.e., the shape of the figure of merit space is parabolic in N-dimensions near local minima. The figure of merit for the control factor optimization is the number of iterations to achieve a solution in comparison to the same case using the standard control factors. This optimization is done for parabolic wells of order N = 2 to 15. In this study it is shown that with the correct choice of the control factors, one can achieve up to a 35% improvement in convergence. Techniques using gradient weighting and the inclusion of additional control factors are proposed.