Measurement error in black dot optimization routine

yuki.moritani noticed that the solutions of black dots optimizations are different between February and January run (even though the setup is the same). This is not unexpected - the black dot optimization algorithm has some inherit uncertainty because the initial simplex for Nelder-Mead is generated from random distribution.
The goal is to estimate what is the distributions of the results between runs that operate on the same dataset. As Robert states in the comments, we need to estimate if these differences are significant. Depending on the result, it might be needed to modify/improve the algorithm.