A comparison of Evolutionary Operation and Simplex for process improvement

• Simplex is highly susceptible to changes in the perturbation size (factorstep).
• EVOP is more robust against noise, especially in higher dimensions.
• Essential parameter in every optimization is the appropriate selection of the perturbation size (factorstep).
• Simplex is preferred when dealing with deterministic or low-noise systems.

Two sequential improvement methods, Evolutionary Operation (EVOP) and Simplex, are compared in a computer simulation study. The EVOP procedure uses underlying statistical models, while the Simplex procedure is based on heuristical rules. On a simple quadratic model the effect of dimensionality, perturbation size – called factorstep in this research – and noise-level on both methods is investigated. Simplex performs quite well in cases with a low level of noise but becomes very unreliable in cases of higher noise and small factorsteps. EVOP is more robust against noise, most noticeable in higher dimensions, but the number of measurements for every step becomes prohibitive with increasing dimensionality.