Agent-based transport models have become increasingly sophisticated tools for understanding mobility patterns and evaluating transportation policies. However, the process of creating these models remains complex and labor-intensive, often requiring extensive manual calibration and region-specific adaptations that limit their transferability across different urban contexts. This dissertation addresses these challenges through methodological contributions spanning two complementary research directions: the integration and calibration of advanced choice modeling techniques, and the development of modular, data-driven approaches for constructing individual model components. One contribution is the development of automated calibration methods that align simulated mode shares with observed data, including segment-based calibration for multiple demographic groups and distance-based approaches using piecewise linear utility functions. Another contribution is the integration of mixed logit models that enables representation of taste heterogeneity, addressing important gaps in current agent-based modeling practice. This approach demonstrates substantial improvements in log-likelihood and produces more realistic behavioral responses to policy changes. Additionally, informed search strategies for co-evolutionary algorithms are developed, achieving faster convergence by incorporating utility estimation and balancing exploration with exploitation. The second part of the thesis addresses the fundamental challenge of creating transferable and reproducible agent-based transport models through systematic, data-driven approaches. Novel methods are presented for estimating essential road infrastructure parameters, including free flow speed estimation through microscopic simulation and machine learning. A comprehensive methodology for extracting and enriching activity facilities from open data sources is developed, incorporating facility attractiveness estimation based on anonymized GPS visitation data. These methods are demonstrated through the creation of the Open Berlin Scenario, which shows substantial improvements compared to the previous model version across multiple validation dimensions. The mean absolute error in mode shares decreased, trip distance distributions showed better alignment with reference data, and traffic volumes and travel speeds demonstrated good agreement with observed data. The Open Berlin Scenario can be used as a benchmark for evaluating new methodological approaches and provides a foundation for testing novel modeling techniques and transport policies in a realistic setting. The modular design of the developed approaches has demonstrated strong transferability across diverse regions, from rural areas to large metropolitan regions in Germany, Mexico, and Japan. The research contributes to more informed decision-making in urban planning and transportation policy by providing robust, automated methodologies that reduce manual effort while improving model accuracy through systematic calibration against real-world observations.
Christian Rakow (Thu,) studied this question.