ExpensiveOptimBenchmark VS parmoo

For the problems with continous decision variables it is not trivial to come up with faster approaches on a modern many-core CPU. But even with discrete input (scheduling and planning) new continous optimizers can compete. The trick is to utilize parallel optimization runs and numba to perform around 1E6 fitness evaluations each second. Advantage is that it is much easier to create a fitness function than for instance to implement incremental score calculation in Optaplanner. And it is more flexible if you have to handle non-standard problems. For very expensive optimizations (like https://github.com/AlgTUDelft/ExpensiveOptimBenchmark) parallelization of fitness evaluation is more important than to use surrogate models.

DTLZ problems are the standard that most papers compare on. I have an implementation of most of DTLZ problems in my package (https://github.com/parmoo/parmoo). You can also find them in pymoo, and most other packages

Then another modified DTLZ set in Python so that the solutions will have a configurable location here (https://github.com/parmoo/parmoo)

I want to plug my own tool, but it is not exactly BO, but could be used for that: - https://github.com/parmoo/parmoo

I maintain an open-source Python library for doing generic surrogate-model-based global optimization (BO being the most common flavor thereof) for the Dept of Energy. https://github.com/parmoo/parmoo