[wenc]

Fascinating! I skimmed the details in the paper here [1] and my impression is that it looks pretty solid. It uses Gutmann's RBFs as surrogate structures. It also benchmarks pretty well against stochastic/metaheuristic algorithms [2].

However most examples in the paper were somewhat small and the problem type is restricted to MINLPs with box constraints (which is still tremendously useful, especially for hyperparameter optimization in simulations).

Just curious, what's the largest problem size you've managed to solve? (no. of constraints, binaries/continuous vars)

[1] http://www.optimization-online.org/DB_FILE/2014/09/4538.pdf

[2] https://www.sciencedirect.com/science/article/pii/S228843001...

[RandomWorker]

Hi paper, is in the works, we designed a bilevel optimization with rbfopt on the upper level and a linear problem on the lower level. Number of constraints are in the 1000s for the linear problem. The inputs to the lower level are 17, and a single output object. We find 400 evaluations of the lower level model for the most important object we will converge to a global minimum. I test ga’s swarm couple of other they didn’t even see the solution found by rbfopt in 800 evaluations.