Machine learning methods in microscopic pedestrian and evacuation dynamics simulation: a comparative study

The modeling and simulation of pedestrian and evacuation dynamics provides essential insights for the field of crowd safety against the background of population increasing and regional development. With the superior performance of machine learning methods demonstrated in pedestrian modeling, varying data encoding schemes and machine learning algorithms were investigated and lack of comparative analysis. Hence, this study analyzes machine learning methods for simulating microscopic pedestrian and evacuation dynamics. The motion interaction field along with a data extraction rule that standardizes input lengths for learning-based models is proposed. Two typical algorithms, Classification and Regression Trees (CART) and Artificial Neural Networks (ANN), are employed for model training and comparison. The fitting performance is evaluated using mean absolute error of velocity, revealing that the CART-based model outperforms the ANN-based model in stability and lower error rates, particularly in varying local density ranges. Dynamics tests are further performed to examine the two models’ robustness against inherent error. The results indicate that the CART-based model struggles under high-density conditions due to the split-based structure. In contrast, the ANN-based model demonstrates superior non-linear fitting ability, allowing for better reproduction of pedestrian dynamics at relatively higher densities. Moreover, the Wasserstein Distance with Sinkhorn iteration is used to quantify model performance in terms of flow-density fundamental diagrams, highlighting the advantages of learning-based approaches over traditional social force model. This research has significant implications for the field of building and civil engineering, as insights from comparative analysis of two typical machine learning algorithms and the establishment of motion interaction field can inform the progress of learning-based pedestrian and evacuation dynamics simulation. The study presented underscores the transformative potential of machine learning methods in simulating pedestrian dynamics and suggests future research directions to enhance robustness and applicability across diverse scenarios of learning-based methods in microscopic pedestrian and evacuation dynamics simulation.