Deep learning approaches for synthetic travel demand modeling with a comparative analysis of methods

Traditionally, transport planning has focused on large flows between spatial regions and gave insights on the need for constructing new road or rail infrastructure. Today, the focus has shifted towards making efficient use of the existing infrastructure with limited changes to the built environment. Such an efficient use requires an intelligent and dynamic matching between demand (travelers and their mobility needs) and supply (transport offers and services). Like the developments in the real world, tools for transport planning are moving from flow-based zonal analysis tools to agent-based simulations in which the dynamic interactions between demand and supply can be modeled and studied in detail. A precondition for performing such simulations are synthetic demand data sets which describe in detail households, persons, and their daily movements in a given territory. To allow for analyses on a high level of detail, sociodemographic attributes of the households and persons need to be congruent with reality, as well as the movements that should follow realistic daily travel patterns. Those can then be used in a detailed transport simulation to explore which services are the most ecological, comfortable and/or cost-efficient in a given territorial context. Applications reach from the assessment of on-demand versus fixed-line services for first/last mile access to the rail infrastructure (Leffler et al., 2021), assessing and mitigating disruptions (Leng and Corman, 2020), or quantifying externalities such as congestion and emissions (Kaddoura et al., 2020) or noise (Le Bescond et al., 2021). With strong links to the transport applications, synthetic populations have recently been used to study household energy consumption patterns (Panos and Margelou, 2019) or model epidemics (Nagel et al., 2021). Consequently, two steps have to be achieved: population synthesis where the goal is to create a synthetic population that should be representative of the real one, and demand synthesis which refers to the process of creating a synthetic travels that are as close as possible to the real demand. Various approaches for population (households and persons) and demand (movements) synthesis have been proposed. For population synthesis, approaches based on sample weighting are most common (Yameogo et al., 2021), while alternative methods based on Bayesian networks (Sun and Erath, 2015), Hierarchichal Mixture Models (Sun et al., 2018), Hidden Markov Models (Saadi et al., 2016b) or Deep Neural Networks (Borysov et al., 2019) have been proposed in recent years. To generate the daily movements of a synthetic population, statistical matching approaches are common (Namazi-Rad et al., 2017) that are challenged by novel Bayesian Network-based methods (Joubert and de Waal, 2020) or Hidden Markov Models (Saadi et al., 2016a). Other aspects such as location choice (Yoon et al., 2012) and discretionary activity assignment (Hörl and Axhausen, 2021) are even less covered in literature. As rightly pointed out by (Garrido et al., 2020), the problem of curse of dimensionality raises when the number of attributes describing the agents and/or their level of detail becomes

large. Many attribute combinations are missing from the sample data while they exist in the real population. The enrichment of the generation process for synthetic travel demand using other data sources have been investigated in (Matet et al. 2023). The authors explore the use of a time-dependant origin-destination (OD) matrix derived from mobile phone data for the attribution of locations in a synthetic population. Approaches based on deep generative latent models (Garrido et al., 2020) such as Generative adversarial networks (GANs) and variational autoencoders (VAEs), have also been proposed to tackle the issue related to the curse of dimensionality and generate synthetic data sets.

To date, there is an evident gap in literature on comparing those competing or complementary approaches with each other and consistent criteria for comparison are missing. Moreover, only few deep learning approaches have been proposed in the literature to tackle synthetic travel demand modeling. The thesis will put in place a comparative analysis of existing population synthesis methods and propose, where applicable, new extensions to the existing methods or combinations thereof. We aim to go beyond evaluating what has been done in literature to designing and proposing new AI based approaches. In order to allow for a comparative assessment of synthetic travel demand generation methods, (Hörl and Balac, 2021) have developed a consistent pipeline from raw, publicly available open data sets to final synthetic travel demand data sets for France. The approach, hence, provides an ideal test bed for benchmarking and comparing existing methods along the generation process. The comparison should be carried out in a horizontal (various algorithms in either population synthesis, location assignment, trip assignment, etc.) and vertical fashion (reconfiguring individual processing steps along an end-to-end generation pipeline). To that end, evaluation criteria need to be developed and consolidated regarding representativeness of the synthetic data and variability (vs. overfitting) towards future and policy scenarios. Furthermore, a special focus should be put on formalizing the flow of information through the modeling pipeline, allowing to consistently assess which correlations between environmental, infrastructural, social, and travel characteristics are expected to have an impact on each other. The thesis project will follow a sequence of work activities:

• Literature review on existing algorithms along the synthetic travel demand pipeline, including classification of the existing approaches and identification of algorithmic gaps.
• Development of a benchmarking pipeline for synthetic travel demand. Integration of the required functionality to track correlations between the inputs and outputs of applied models. Integration of a set of relevant algorithms and methods for comparison, especially taking into account methods for the processing of mobility traces.
• Definition of key performance indicators on the quality and responsiveness of synthetic travel demand data sets, also taking into account the correlation structure of the generated data set.
• Development of AI-based approaches for travel population synthesis methods using multisource data (survey, mobile phone data, public transport data).
• Benchmarking of the implemented approaches in terms of the defined indicators and in terms of data granularity. Provision of a roadmap for future developments on synthetic travel demand modeling, taking into account the obtained results and insights.

International Mobility As part of this project, we are planning to collaborate with Prof. Lijun Sun, and arrange for a PhD student to stay at McGill University for a 3-month research mobility. A collaboration with Prof. Klaus Bogenberger (Technical University of Munich) is also possible during the thesis.

References Borysov, S.S., Rich, J., Pereira, F.C., 2019. How to generate micro-agents? A deep generative modeling approach to population synthesis. Transp. Res. Part C Emerg. Technol. 106, 73–97. https://doi.org/10.1016/j.trc.2019.07.006

Garrido S., Borysov S.S., Pereira F.C., Rich J. Prediction of rare feature combinations in population synthesis: Application of deep generative modelling. Transp. Res. C, 120 (2020), https://doi.org/10.1016/j.trc.2020.102787

Hörl, S., Axhausen, K.W., 2021. Relaxation–discretization algorithm for spatially constrained secondary location assignment. Transp. Transp. Sci. 1–20. https://doi.org/10.1080/23249935.2021.1982068

Hörl, S., Balac, M., 2021. Synthetic population and travel demand for Paris and Île-de-France based on open and publicly available data. Transp. Res. Part C Emerg. Technol. 130, 103291. https://doi.org/10.1016/j.trc.2021.103291

Joubert, J.W., de Waal, A., 2020. Activity-based travel demand generation using Bayesian networks. Transp. Res. Part C Emerg. Technol. 120, 102804. https://doi.org/10.1016/j.trc.2020.102804

Kaddoura, I., Bischoff, J., Nagel, K., 2020. Towards welfare optimal operation of innovative mobility concepts: External cost pricing in a world of shared autonomous vehicles. Transp. Res. Part Policy Pract. 136, 48–63. https://doi.org/10.1016/j.tra.2020.03.032

Le Bescond, V., Can, A., Aumond, P., Gastineau, P., 2021. Open-source modeling chain for the dynamic assessment of road traffic noise exposure. Transp. Res. Part Transp. Environ. 94, 102793. https://doi.org/10.1016/j.trd.2021.102793

Leffler, D., Burghout, W., Jenelius, E., Cats, O., 2021. Simulation of fixed versus on-demand station-based feeder operations. Transp. Res. Part C Emerg. Technol. 132, 103401. https://doi.org/10.1016/j.trc.2021.103401

Leng, N., Corman, F., 2020. The role of information availability to passengers in public transport disruptions: An agent-based simulation approach. Transp. Res. Part Policy Pract. 133, 214–236. https://doi.org/10.1016/j.tra.2020.01.007

Matet, B., Côme, E., Furno, A., Hörl, S., Oukhellou, L., 2023. Use of Origin-Destination data for calibration and spatialization of synthetic travel demand. 11th Symposium of the European Association for Research in Transportation (hEART 2023), Zurich, Swiss.

Nagel, K., Rakow, C., Müller, S.A., 2021. Realistic agent-based simulation of infection dynamics and percolation. Phys. Stat. Mech. Its Appl. 584, 126322. https://doi.org/10.1016/j.physa.2021.126322

Namazi-Rad, M.-R., Tanton, R., Steel, D., Mokhtarian, P., Das, S., 2017. An unconstrained statistical matching algorithm for combining individual and household level geo-specific census and survey data. Comput. Environ. Urban Syst. 63, 3–14. https://doi.org/10.1016/j.compenvurbsys.2016.11.003

Panos, E., Margelou, S., 2019. Long-Term Solar Photovoltaics Penetration in Single- and Two-Family Houses in Switzerland. Energies 12, 2460. https://doi.org/10.3390/en12132460

Saadi, I., Mustafa, A., Teller, J., Cools, M., 2016a. Forecasting travel behavior using Markov Chains-based approaches. Transp. Res. Part C Emerg. Technol. 69, 402–417. https://doi.org/10.1016/j.trc.2016.06.020

Saadi, I., Mustafa, A., Teller, J., Farooq, B., Cools, M., 2016b. Hidden Markov Model-based population synthesis. Transp. Res. Part B Methodol. 90, 1–21. https://doi.org/10.1016/j.trb.2016.04.007

Sun, L., Erath, A., 2015. A Bayesian network approach for population synthesis. Transp. Res. Part C Emerg. Technol. 61, 49–62. https://doi.org/10.1016/j.trc.2015.10.010

Sun, L., Erath, A., Cai, M., 2018. A hierarchical mixture modeling framework for population synthesis. Transp. Res. Part B Methodol. 114, 199–212. https://doi.org/10.1016/j.trb.2018.06.002

Yameogo, B.F., Vandanjon, P.-O., Gastineau, P., Hankach, P., 2021. Generating a Two-Layered Synthetic Population for French Municipalities: Results and Evaluation of Four Synthetic Reconstruction Methods. J. Artif. Soc. Soc. Simul. 24, 5. https://doi.org/10.18564/jasss.4482

Yoon, S.Y., Deutsch, K., Chen, Y., Goulias, K.G., 2012. Feasibility of using time–space prism to represent available opportunities and choice sets for destination choice models in the context of dynamic urban environments. Transportation 39, 807–823. https://doi.org/10.1007/s11116-012-9407-8