Deep graph learning and bagging-based metamodel for efficient simulation-based calibration of large-scale non-euclidean road network attributes

As the major supply input for traffic simulations, road network attribute calibration is essential but challenging, due to the high dimensionality in the non-Euclidean solution space. This paper presents an efficient simulation-based optimization (SBO) framework for large-scale non-Euclidean decision variables in metropolitan networks. The graph learning technique is introduced to capture the correlations among decision variables and cluster them into homogeneous groups, enabling cluster-specific optimization that reduces the search space. A bagging-based physical metamodel is further developed for optimization acceleration by replacing the time-consuming simulator with multiple simplified traffic models in gradient estimation. Each traffic model is built on randomly sampled travel demands to guarantee computational efficiency. The framework is evaluated by calibrating the capacities and speed limits of 8246 segments in a real-world metropolitan road network within limited iterations. Both the clustering model and the metamodel outperform their counterparts, achieving lower simulation error and reduced computational time.