Biobjective robust simulation-based optimization for unconstrained problems

We propose a biobjective robust simulation-based optimization (BORSO) method to solve unconstrained problems involving implementation errors and parameter perturbations. We adopt the notion that a solution is robust efficient (RE) if the region that dominates its worst-case realizations of the biobjectives under uncertainty does not contain (all) the worst-case realizations of the biobjectives of any other solution under uncertainty. Our algorithm aims to efficiently find a set of RE solutions through a series of function evaluations or simulations. First, we design surrogate-model guided search mechanisms for the worst-case neighbors of the current iterate. Subsequently, we determine the iteration distance along an effective local move direction, which excludes the worst-case neighbors from the neighborhood of the new iterate. Depending on the practical need for solution diversity, multiple initial solutions can be specified in our algorithm, and the final iterates of these solutions generate a set of RE solutions. The test results of a synthetic biobjective nonconvex optimization problem show the effectiveness of the BORSO method and its superior performance against a sampling-based robust multiobjective optimization solver. Furthermore, when the proposed algorithm is applied to a real-world biobjective traffic signal timing problem, satisfactory solutions can be obtained under a limited computational budget. These results indicate that the proposed BORSO method can solve unconstrained biobjective simulation-based optimization problems under uncertainties.