Network optimization with an embedded traffic assignment (TA) model is at the heart of transportation planning and operations. This problem becomes intractable for large-scale multimodal networks with high dimensional decision variables due to nonlinearity, high computation time, and lack of closed-form expressions. The problem intensifies significantly with modern agent-based microsimulation models that might take days for a single evaluation. General-purpose algorithms, heuristics, and problem-specific bi-level formulations have been proposed in the past to solve small problems for demonstration purposes and certain large-scale problems. Research gap, however, exists in developing efficient solution methods for high-dimensional, large-scale multimodal problems. In this thesis, we cont...[ Read more ]

Network optimization with an embedded traffic assignment (TA) model is at the heart of transportation planning and operations. This problem becomes intractable for large-scale multimodal networks with high dimensional decision variables due to nonlinearity, high computation time, and lack of closed-form expressions. The problem intensifies significantly with modern agent-based microsimulation models that might take days for a single evaluation. General-purpose algorithms, heuristics, and problem-specific bi-level formulations have been proposed in the past to solve small problems for demonstration purposes and certain large-scale problems. Research gap, however, exists in developing efficient solution methods for high-dimensional, large-scale multimodal problems. In this thesis, we contribute to mitigating this research gap by addressing three major challenges. Firstly, we address the problem of simulator inefficiency by developing a metamodel-based optimization framework for computationally expensive TA models. Multiple gradient-based metamodels and simplified traffic model embedded metamodels were presented as part of this framework, via which we successfully calibrated an agent-based, multi-modal traffic microsimulation model of Hong Kong Island (HKI). We also develop and test a kriging metamodel for dynamic network loading with high parallelization potentials and improved efficiency. Secondly, we address the problem of high dimensionality by developing an efficient TA gradient estimation technique called iterative backpropagation (IB). IB produces gradients for all equilibrium TA outputs using only a single TA equilibrium evaluation, irrespective of the dimension of the problem, thus making it both versatile and highly efficient. Thirdly, we address the problem of added complexity due to agent-based formulations. We propose an agent-based TA model with optional transit mobility packages and optimize the system for the HKI multimodal network with an extended IB algorithm. The methodologies proposed in this study have numerous applicability in transportation planning and operations, especially in enhancing the efficiency and profitability of the underlying transportation systems.