Enhancing activity-based models: dynamic scheduling and trip chain modeling

Given the fundamental role of mobility and transportation in human prosperity, intelligent implementation of transportation systems is essential to eliminate, or at least minimize, the negative externalities associated with them. One approach to reducing these external costs involves comprehensively understanding the root causes of the challenges and subsequently addressing them. Solutions to these challenges may be technical, such as the development of cleaner and more efficient fuels, as well as nontechnical, such as the implementation of policy instruments. The design of these instruments requires specific knowledge, which can be gathered through travel demand modeling. Although travel demand models have advanced significantly, the phenomena of urbanization and the urgent threat of global warming present novel policy questions that require more specialized and comprehensive travel demand models. This dissertation significantly enriches the academic discourse on travel demand analysis, thereby promoting the advance of the state-of-the-art Activity-Based Model (ABM)s. It addresses two key aspects of travel demand modeling: the temporal dimension of day-to-day variability in the participation of discretionary activities and the travel mode choice dimension. The variability of day-to-day discrete activities can be examined through the application of the inter-episode duration (IED)s. To achieve the aims of this dissertation, it commences with a comprehensive literature review, showcasing the existing knowledge surrounding IED and travel mode choice modeling. An analysis of the literature and the current state of the art ABMs reveals a strong consensus on the necessity of ABMs with a multi-day forecasting scope. Consequently, inspired by methodologies from other domains, there has been a growing focus on the use of hazard models to investigate IEDs. The literature review further illuminates the domain of travel mode choice modeling. In this context, a critical evaluation of various methodologies, including discrete choice models and machine learning classifiers, is conducted. In addition, the significance and impact of predictor variables are systematically analyzed. Insights from the literature review are then incorporated into the methodologies. Two main data sources are used. Firstly, to analyze IEDs, a large number of observations are required to properly estimate the proposed models. The Nederlands Verplaatsingspanel (NVP) dataset, a vast Global Positioning System (GPS) dataset containing activity-travel schedules over multiple days or months even, is used in the IED studies. Secondly, in the travel mode choice modeling frameworks, the data in the Mobiliteitspanel Nederland (MPN) panel are used. The first analysis studies IEDs for discretionary activities grocery shopping (524,439 observations by 11,024 respondents), non-grocery shopping (401,065 observations by 9,218 respondents) and leisure activities (128,818 observations by 4,836 respondents), using the NVP data, collected throughout the country of the Netherlands. Exponential, Erlang-????, and exponential-Erlang-2 mixture models have been employed, in which individuals are latently classified as randoms or regulars, depending on whether a better fit is observed with the exponential or the Erlang-2 segment, respectively. The results indicate that the proportion of individuals engaging in grocery activities in a random manner is the smallest, whereas it is the largest for leisure activities. Furthermore, the hazards are adapted to include both time-varying, such as weather conditions and day of the week, as well as time-independent covariates, such as household and personal characteristics, and the results reveal that particularly harsh weather conditions significantly affect the underlying hazards. Unobserved heterogeneity is accounted for in both a nonparametric and a parametric fashion, to address concerns that the use of parametric forms to account for unobserved heterogeneity is not always warranted. The results suggest that gamma heterogeneity works well for the exponential and mixture models, while nonparametric heterogeneity works better for the Erlang-2 model. A subsequent analysis extends the first analysis in the sense that here, too, IEDs are studied. However, more flexible baseline hazards—the shifted Weibull and the shifted log-logistic (SLL)—are proposed. Contrary to the exponential and Erlang-2 models, the Weibull and SLL models allow for monotonically increasing and unimodal hazards, respectively. The estimated parameters indicate that for all three activities monotonically increasing and unimodal hazards are found, indicating that for the former, the likelihood that an activity will happen increases over time. For the latter, a peak in hazard can be observed around two days. Time-varying and time-independent characteristics are again included, and the latter are extended to also include land use characteristics, such as distances to amenities. Unobserved heterogeneity is included in a nonparametric fashion, leading to improved model fits. Finally, a mixture (Weibull-SLL) is proposed and briefly compared to a shifted generalized gamma model. The third analysis shifts away from the IED analysis and focuses on travel mode choice analysis. More specifically, it proposes a multi-level model (MLM) framework in which the fixed effects are estimated by a chi-square automatic interaction detection (CHAID) tree or a Gradient-Boosted Decision Tree (GBDT) model. This framework leverages the strengths of nonparametric machine learning classifiers while estimating panel effects using the MPN panel dataset. Multiple random intercept models are proposed, explaining variance on different levels, resulting in increased predictive performance with increasing model complexity. A wide range of predictor variables is considered that range from household and individual characteristics, weather conditions, land use characteristics, and trip characteristics. The fourth and final analysis continues in the travel mode choice modeling domain and disentangles travel mode choices for multimodal trips, for trips consisting of a maximum of three trip legs. A classifier chain is proposed to successively predict each of the choices in such trips, considering the predictions of previous choices as input for the current prediction. Studying the outcomes and the importance of the variables reveals the (partial) label dependence with respect to predictions from previous steps. Other important predictors are trip characteristics and for multimodal trips, distances to Public Transport (PT) services such as the train station. Finally, this dissertation concludes by highlighting both the scientific and the societal contributions of the presented work. The relevance of pushing the frontiers of travel demand modeling, and more specifically ABMs, is again emphasized by the increasing urgency of urban challenges, such as inequality, climate change, and urbanization.