Learning-Based Memory-Guided Trust-Region Bayesian Optimization (MG-TuRBO)

Abstract

Traffic digital twins provide realistic system-level testbeds for intelligent transportation system applications. Their value depends on how well they reproduce observed traffic conditions, making calibration against data essential. In practice, traffic digital-twin calibration is a black-box optimization problem in which each simulator evaluation is expensive and the parameter space may be high-dimensional. To address this challenge, we propose two complementary contributions within a Bayesian optimization framework: Memory-Guided Trust-Region Bayesian Optimization (MG-TuRBO) and an adaptive acquisition strategy. MG-TuRBO improves restart decisions by using search memory to identify basins that are both promising and under-explored, while the adaptive acquisition strategy, a separate contribution, uses the remaining evaluation budget to balance exploration and exploitation during candidate selection. We demonstrate the method on two real-world traffic simulation problems with 14-dimensional (14D) and 84-dimensional (84D) calibration spaces and compare it against a genetic algorithm (GA), classical Bayesian optimization (BO), Trust-Region Bayesian Optimization (TuRBO), and Multi-TuRBO. Bayesian optimization methods reach target calibration quality much faster than GA in 14D. In 84D, MG-TuRBO Adaptive significantly outperforms Multi-TuRBO, GA, and BO and is statistically indistinguishable from TuRBO Adaptive, which has the lowest final median GEH (3.07 vs. 3.10, $p=1.00$ ). MG-TuRBO Adaptive shows the strongest observed target-reaching behavior for the GEH < 4 and GEH < 3.5 targets and the strongest observed restart effectiveness. MG-TuRBO is promising for transportation digital-twin calibration and potentially applicable to other expensive high-dimensional black-box optimization problems, subject to broader validation.