Traffic simulation and digital-twin calibration is a challenging optimization problem with a limited simulation budget. Each trial requires an expensive simulation run, and the relationship between calibration inputs and model error is often nonconvex, and noisy. The problem becomes more difficult as the number of calibration parameters increases. We compare a commonly used automatic calibration method, a genetic algorithm (GA), with Bayesian optimization methods (BOMs): classical Bayesian optimization (BO), Trust-Region BO (TuRBO), Multi-TuRBO, and a proposed Memory-Guided TuRBO (MG-TuRBO) method. We compare performance on 2 real-world traffic simulation calibration problems with 14 and 84 decision variables, representing lower- and higher-dimensional (14D and 84D) settings. For BOMs, we study two acquisition strategies, Thompson sampling and a novel adaptive strategy. We evaluate performance using final calibration quality, convergence behavior, and consistency across runs. The results show that BOMs reach good calibration targets much faster than GA in the lower-D problem. MG-TuRBO performs comparably in our 14D setting, it demonstrates noticeable advantages in the 84D problem, particularly when paired with our adaptive strategy. Our results suggest that MG-TuRBO is especially useful for high-D traffic simulation calibration and potentially for high-D problems in general.
Figures & tables
Fig. 1: SUMO models of the calibration networks: (Left) 14D Shallowford Road network (Right) 84D Murfreesboro Road corridor network.
Fig. 2: Optimization-phase convergence for the 14D network.
Fig. 3: PCA-based 2D exploration patterns for the 14D network comparing GA with TuRBO methods under Thompson Sampling. The best for each method is marked with a star.
Fig. 4: Final best GEH distributions for Thompson Sampling and Adaptive acquisition across BOMs, with GA as a shared baseline.
Fig. 5: Optimization-phase convergence for the 84D calibration problem.
Fig. 6: Trust-region search mechanisms in 84D (Adaptive acquisition). TuRBO executes 20 sequential restarts, each time exploring a single region until failure. Multi-TuRBO maintains concurrent regions exploring different basins. MG-TuRBO cycles through 21 local regions, allocating small budgets to systematically sample the landscape.
GERAD & Department of Mathematics and Industrial Engineering, Polytechnique Montr´eal. · GERAD & Department of Management, Universit´e du Qu´ebec `a Trois-Rivi`eres.