Comparative analysis of pedestrian volume models: Agent-based models, machine learning methods and multiple regression analysis

Pedestrian flow distributions can inform planning for walkability and improve understanding of factors that influence pedestrian activity. However, detailed data is rarely available so pedestrian volume models, commonly relying on the Space Syntax framework, are often utilized to predict pedestrian volumes. This study compares the performance and dominant variables of three modelling families – multiple regression analyses, machine learning models, and agent-based models – in Tel Aviv-Yafo, Israel. Using 247 flow observations, optimal models from each family were fitted and validated for 3 separate areas that differ in their urban growth and morphological characteristics, as well for the whole city. Results showed that ensemble-based machine learning models were best for city-wide predictions while agent-based models had an advantage at the local scale of neighborhoods – especially in neighborhoods that did not develop in a self-organized process. Regression analyses fell short for all areas, even when using principal component analysis to reduce multicollinearity and overfitting. These differences are attributed to the relative influence of cognitive-behavioral and structural factors on pedestrian flows: agent-based models outperform statistical models in individual areas, where behavior is captured more accurately using a small set of cognitive-behavioral parameters. Statistical models are dominant in the city-wide context, where structural variables can predict aggregate patterns. This is crucially important when evaluating the distribution of pedestrians in a planned urban environment. Overall, our results indicate that stepwise regression are not sufficient for pedestrian volume modelling, that agent-based models better capture complex interactions between independent variables, and that machine learning models have a strong potential for city-wide pedestrian volume modelling.