A revised simplex search procedure for stochastic simulation response surface optimization

We develop a variant of the Nelder–Mead (NM) simplex search procedure for stochastic simulation optimization that is designed to avoid many of the weaknesses encumbering similar direct-search methods – in particular, excessive sensitivity to starting values, premature termination at a local optimum, lack of robustness against noisy responses, and computational inefficiency. The Revised Simplex Search (RSS) procedure consists of a three-phase application of the NM method in which: (a) the ending values for one phase become the starting values for the next phase; (b) the step size for the initial simplex (respectively, the shrink coefficient) decreases geometrically (respectively, increases linearly) over successive phases; and (c) the final estimated optimum is the best of the ending values for the three phases. To compare RSS versus NM and procedure RS–S9 due to Barton and Ivey, we summarize a simulation study based on four selected performance measures computed for six test problems that include additive white-noise error, with three levels of problem dimensionality and noise variability used in each problem. In the selected test problems, RSS yielded significantly more accurate estimates of the optimum than NM or RS–S9, and both RSS and RS–S9 required roughly four times as many function evaluations as NM.