SCR: A novel surrogate-based global optimization algorithm for constrained black-box problems

This paper proposes a novel optimization algorithm for constrained black-box problems, where the objective function and some constraints are computed by a simulation code. The basic idea of the optimization algorithm, referred to as SCR (Surrogate-CMAESRQLIF), is to (i) build separate Kriging surrogates for the objective function and black-box constraints, (ii) use the global search algorithm CMAES to find the global optimum region of the surrogate, (iii) use the recent algorithm RQLIF to refine the search locally, (iv) use all the points sampled by RQLIF and additional points within the optimal region located by CMAES to update the surrogate model. Tests on 46 constrained and unconstrained test problems show that SCR outperforms the benchmark algorithms in terms of fraction of problems solved, specially at low function evaluations (< 300).