MAiNGO not always converging
I am not sure if this is a bug or something that can be expected normally. I found that the MAiNGO-based minimum of the Gaussian process mean prediction is not always the actual (global) minimum. I some rare cases one can find a better optimum by taking the MAiNGO-based solution as the initial value for a local optimizer. I produced the figure below by modifying example_training_gp_pymelon.py slightly. I use the SLSQP solver in SciPy to search locally around the optimum obtained with MAiNGO. In this particular instance, a better optimum can be found this way.
I am curious if the precision of the MAiNGO software can be influenced in some way.
This is the modified code to reproduce this issue: example_training_gp_pymelon.py