Optimizing an expensive multi-objective building performance problem: Benchmarking model-based optimization algorithms against metaheuristics with and without surrogates

• Popular multi-objective metaheuristic optimization algorithms are benchmarked against state-of-the-art model-based optimizers. • Benchmark is conducted on a real-world building performance simulation problem. • Model-based optimization outperforms metaheuristics both in terms of robustness and achieved hypervolume. • Substituting the simulation model with a machine learning surrogate to speed up metaheuristic optimization improves their performance but does not yield better results than model-based optimization. While simulation-based optimization can effectively find good solutions, the need to simulate hundreds of candidates and consequent long run-times prevent their application in practice. Accurate and fast surrogate models can replace expensive building performance simulations (BPS). Model-based optimization algorithms construct a surrogate during optimization and perform many additional optimization steps quickly. While this strategy has proven effective for expensive single-objective optimization, its performance on multi-objective BPS problems remains understudied. Two questions persist: A) Do model-based multi-objective optimization algorithms outperform metaheuristics and B) How does optimizing on a surrogate model affect the performance of metaheuristic optimization algorithms? Our benchmark results show that the model-based algorithms RBFMOpt and TPE outperform metaheuristics regarding robustness, maximum hypervolume, and the quality of the found Pareto fronts. RBFMOpt yields good solutions within less than 100 function evaluations. Optimizing on surrogate models heavily depends on the surrogates’ ability to estimate precisely but is computationally cheap and allows larger optimization budgets.