Enriching and validating urban-scale agent-based space-syntax pedestrian-volume models across diverse urban morphologies

Omer and Kaplan (2017) introduced an urban-scale agent-based (AB) pedestrian-volume model that encodes space-syntax centrality concepts and land-use attractors into explicit agent route-choice rules, and showed on two Israeli cities that it can out-predict the conventional multiple regression analysis (MRA) benchmark out of sample in a planned/hierarchical street network while matching it in a self-organized near-grid one — but on very small validation samples (8 street segments per test area). The authors themselves framed the unresolved problem: whether enriching the framework with additional network centrality measures (e.g., geographically modified PageRank variants), alternative spatial representations (axial lines, natural streets), and richer agent specifications (population density, transportation-mode and transit access, heterogeneous agent types, diurnal dynamics), and validating it systematically across morphologically diverse urban areas, improves out-of-sample prediction and yields a quantified domain of validity. A same-group follow-up (Wolpert and Omer 2025) has since resolved part of the comparison question within a single city (machine learning best city-wide, AB best locally, regression insufficient), but the agent-enrichment components and the multi-city domain-of-validity question remain open.

Background

Planners need the distribution of pedestrian volumes over entire street networks, but counting pedestrians on every street is infeasible, so models that predict segment-level volumes from network and land-use data are a core urban-analytics tool. The dominant tradition is space syntax: a configurational analysis of the street network represented as street segments, in which each segment receives centrality scores. Choice (also called betweenness) counts how often a segment lies on shortest paths between all origin-destination pairs and proxies through-movement; Integration (a closeness-type measure) proxies to-movement. Shortest paths can be defined by metric distance (geometric length, e.g. via Dijkstra's algorithm), topological distance (fewest turns), or angular distance (minimum cumulative turning angle), and each centrality can be restricted to a radius so that only trips up to a given scale are counted. In the standard approach, log pedestrian counts on a sample of observed segments are regressed on such centrality measures plus on-segment land-use counts by multiple regression analysis (MRA), typically evaluated only in-sample. The source line of work is Lerman, Rofè and Omer (2014) and Omer, Rofè and Lerman (2015), which contrasted pedestrian movement models in pre-modern and modern Israeli neighborhoods. Omer and Kaplan (2017) asked whether an agent-based model (ABM — a simulation of many individual agents whose local decision rules generate aggregate patterns) could replace the regression: agents choose shortest paths by metric, topological, or angular distance within a movement radius, and pick destinations by a land-use attractor weight $\mathrm{LU}_A(S_i)$ proportional to weighted counts of non-residential buildings per segment. The model was implemented by coupling GIS data to NetLogo (an ABM platform) and calibrated by parameter sweeps in BehaviorSpace (NetLogo's built-in batch-experiment tool; radii 500–5000 m, agent-type mixtures, land-use weights) against observed counts in Kfar Saba (312 segments; 17 calibration, 8 validation) and Ashdod (457 segments; 20 calibration, 8 validation). Aggregate simulated flows reproduced the matching Betweenness centralities at Pearson $r=0.93$–$0.99$, calibration fits reached $R^2\approx0.90$–$0.92$ (MRA: 0.96 and 0.93), and on adjacent out-of-sample test areas the AB model matched MRA in the self-organized, near-grid city (Kfar Saba: $R^2=0.593$ vs $0.590$) and clearly beat it in the planned, hierarchical city (Ashdod: $R^2=0.621$ vs $0.282$), the mechanism being that routed agents accumulate through-movement on conduit segments to remote attractors that regression, which sees only local centrality and on-site land use, under-predicts. The authors stress that the validation samples are tiny (8 segments), that no significance tests accompany the comparison, and that the study covers only two cities; they explicitly recommend as next steps: incorporating additional centrality measures such as geographically modified PageRank (a location-weighted variant of Google's PageRank ranking algorithm, introduced for spatial networks by Chin and Wen 2015) and alternative spatial representations such as axial lines (the classical space-syntax representation by longest sight lines) and natural streets (streets stitched from same-name or continuously aligned segments); enriching agents with population-density effects, transportation-mode availability and transit access, heterogeneous agent types with differential movement potentials, and diurnal (time-of-day) dynamics; and systematically testing applicability and robustness across urban areas with diverse growth histories, street patterns, and land-use distributions to quantify the model's domain of validity. Since then, adjacent literatures have grown on parallel tracks: machine-learning pedestrian-volume estimation (reviewed by Lian, Loo and Fan 2024; e.g. Park, No and Kim 2025), diurnal statistical models of pedestrian concentration (Li, Song and Zhou 2020; Jiang, Ma and Li 2022), and cognitive enrichment of pedestrian ABMs (Filomena and colleagues' region, barrier, and landmark route-choice models; Wozniak, Filomena and Wronkowski's route-choice typology) — but none of these has performed the enrichment-plus-systematic-validation program on the AB space-syntax framework itself.

Problem Statement

Does enriching the urban-scale agent-based space-syntax pedestrian-volume framework of Omer and Kaplan (2017) — by incorporating additional network centrality measures (for example, geographically modified PageRank variants) and alternative spatial representations (axial-line and natural-street models in place of, or alongside, segment maps), and by enriching the agent specification and input data with population-density effects, transportation-mode availability and access to public transit, heterogeneous agent types with differential movement potentials, and temporal dynamics capturing time-of-day variation in movement distributions — improve out-of-sample prediction of pedestrian volumes across a set of urban areas with diverse morphological characteristics (planned/hierarchical versus self-organized/grid street patterns, different growth histories and land-use distributions), relative to the calibrated multiple regression benchmark (and, under current practice, machine-learning baselines)? The question comprises three parts: (a) implementing and calibrating such enriched agent-based models on observed pedestrian counts with the same empirical footing as the regression benchmarks; (b) validating them out of sample in multiple, morphologically diverse urban areas (not only adjacent test areas within one or two cities, and with validation samples large enough to support the claimed differences); and (c) quantifying the framework's domain of validity — for which urban morphologies, movement scales, and time periods each model family (enriched AB, regression, machine learning) predicts best, and why. The question does not presuppose improvement: a rigorous systematic negative result (specified enrichments fail to improve out-of-sample prediction) resolves it on the same footing as a demonstrated improvement. The enrichment components named above define the scope of the program rather than a conjunctive checklist whose full success is required: the components may be evaluated individually and need not all succeed, with each evaluated component standing on the same empirical footing as the benchmarks. The program may be implemented, calibrated, and validated on any publicly available or otherwise accessible observed segment-level pedestrian-count, street-network, and land-use datasets covering the morphologically distinct urban areas studied, provided it meets the same empirical footing as the regression benchmarks in (a)-(b) and does not substitute undocumented or fabricated observations.

References

1. Itzhak Omer, Nir Kaplan (2017), 'Using space syntax and agent-based approaches for modeling pedestrian volume at the urban scale', Computers, Environment and Urban Systems 64:57–67. DOI: 10.1016/j.compenvurbsys.2017.01.007. https://doi.org/10.1016/j.compenvurbsys.2017.01.007
2. Lior Wolpert, Itzhak Omer (2025), 'Comparative analysis of pedestrian volume models: Agent-based models, machine learning methods and multiple regression analysis', Computers, Environment and Urban Systems 117:102238. DOI: 10.1016/j.compenvurbsys.2024.102238. https://doi.org/10.1016/j.compenvurbsys.2024.102238
3. Yoav Lerman, Yodan Rofè, Itzhak Omer (2014), 'Using Space Syntax to Model Pedestrian Movement in Urban Transportation Planning', Geographical Analysis 46(4):392–410. DOI: 10.1111/gean.12063. https://doi.org/10.1111/gean.12063
4. Itzhak Omer, Yodan Rofè, Yoav Lerman (2015), 'The impact of planning on pedestrian movement: contrasting pedestrian movement models in pre-modern and modern neighborhoods in Israel', International Journal of Geographical Information Science 29(12):2121–2142. DOI: 10.1080/13658816.2015.1063638. https://doi.org/10.1080/13658816.2015.1063638
5. Wei-Chien-Benny Chin, Tzai-Hung Wen (2015), 'Geographically Modified PageRank Algorithms: Identifying the Spatial Concentration of Human Movement in a Geospatial Network', PLOS ONE 10(10):e0139509. DOI: 10.1371/journal.pone.0139509. https://doi.org/10.1371/journal.pone.0139509
6. Shuguang Li, Dong Song, Qilong Zhou (2020), 'Estimating Dynamic Distribution Condition of Pedestrian Concentration on an Urban Scale', Journal of Urban Planning and Development 146(4). DOI: 10.1061/(ASCE)UP.1943-5444.0000626. https://doi.org/10.1061/(ASCE)UP.1943-5444.0000626
7. Gabriele Filomena, Ed Manley, Judith A. Verstegen (2020), 'Perception of urban subdivisions in pedestrian movement simulation', PLOS ONE 15(12):e0244099. DOI: 10.1371/journal.pone.0244099. https://doi.org/10.1371/journal.pone.0244099
8. Gabriele Filomena, Judith A. Verstegen (2021), 'Modelling the effect of landmarks on pedestrian dynamics in urban environments', Computers, Environment and Urban Systems 86:101573. DOI: 10.1016/j.compenvurbsys.2020.101573. https://doi.org/10.1016/j.compenvurbsys.2020.101573
9. Chen Yang, Zhu Qian (2023), 'Street network or functional attractors? Capturing pedestrian movement patterns and urban form with the integration of space syntax and MCDA', Urban Design International 28(1):3–18. DOI: 10.1057/s41289-022-00178-w. https://doi.org/10.1057/s41289-022-00178-w
10. Gabriele Filomena, Lia Kirsch, Angela Schwering, Judith A. Verstegen (2022), 'Empirical characterisation of agents' spatial behaviour in pedestrian movement simulation', Journal of Environmental Psychology 82:101807. DOI: 10.1016/j.jenvp.2022.101807. https://doi.org/10.1016/j.jenvp.2022.101807
11. Feifeng Jiang, Jun Ma, Zheng Li (2022), 'Pedestrian volume prediction with high spatiotemporal granularity in urban areas by the enhanced learning model', Sustainable Cities and Society 79:103653. DOI: 10.1016/j.scs.2021.103653. https://doi.org/10.1016/j.scs.2021.103653
12. David Bowater, Emmanuel Stefanakis (2023), 'Extending the Adapted PageRank Algorithm centrality model for urban street networks using non-local random walks', Applied Mathematics and Computation 446:127888. DOI: 10.1016/j.amc.2023.127888. https://doi.org/10.1016/j.amc.2023.127888
13. Ting Lian, Becky P. Y. Loo, Zhuangyuan Fan (2024), 'Advances in estimating pedestrian measures through artificial intelligence: From data sources, computer vision, video analytics to the prediction of crash frequency', Computers, Environment and Urban Systems 107:102057. DOI: 10.1016/j.compenvurbsys.2023.102057. https://doi.org/10.1016/j.compenvurbsys.2023.102057
14. Chulwoong Park, Wonjun No, Youngchul Kim (2025), 'A probabilistic AI-based pedestrian volume estimation model for street-level urban management', Journal of Transport Geography 129:104437. DOI: 10.1016/j.jtrangeo.2025.104437. https://doi.org/10.1016/j.jtrangeo.2025.104437
15. Marcin Wozniak, Gabriele Filomena, Adam Wronkowski (2025), 'What's your type? A taxonomy of pedestrian route choice behaviour in cities', Transportation Research Part F: Traffic Psychology and Behaviour 109:1257–1274. DOI: 10.1016/j.trf.2025.01.012. https://doi.org/10.1016/j.trf.2025.01.012
16. Fatemeh Nourmohammadi, Taha Rashidi, Meead Saberi (2025), 'Spatial Transferability of Pedestrian Trip Generation Models', SSRN preprint. DOI: 10.2139/ssrn.5105541. https://doi.org/10.2139/ssrn.5105541