The matching radius, defined as the maximum pick-up distance within which waiting riders and idle drivers can be matched, is a critical variable in ride-hailing systems. Optimizing the matching radius can significantly enhance system performance, but determining its optimal value is challenging due to the dynamic nature of ride-hailing environments. The matching radius should adapt to spatial and temporal variations, as well as to real-time fluctuations in supply and demand. To address this challenge, this paper proposes a dual-reply-buffer deep reinforcement learning method for dynamic matching radius optimization. By modeling the matching radius optimization problem as a Markov decision process, the method trains a policy network to adaptively adjust the matching radius in response to changing conditions in the ride-hailing system, thereby improving efficiency and service quality. We validate our method using real-world ride-hailing data from Austin, Texas. Experimental results show that the proposed method outperforms baseline approaches, achieving higher matching rates, shorter average pick-up distances, and better driver utilization across different scenarios.

Crowd management plays a vital role in urban planning and emergency response. Accurate crowd prediction is important for venue operators to respond effectively to adverse crowd dynamics during large gatherings. Although many studies have tried to predict crowd densities or movement dynamics with data-driven predictive models, their validation is often limited to data within the same scenario. As a result, the predictability of the data-driven model in unseen scenarios, such as evacuation scenarios, remains unknown due to the challenges of collecting out-of-distribution data regarding emergency conditions. To address this problem, we present an evaluation pipeline to evaluate different kinds of data-driven models. A method is proposed to generate realistic scenarios by simulation and collect synthetic data from these scenarios to acquire a comprehensive dataset. With these synthetic data, we evaluated different predictive models, from traditional machine learning methods to deep learning time-series prediction models, to explore their generalizability. Furthermore, we propose a weighted average metric, which is better suited to determine the performance of forecasting algorithms under adverse conditions. Through extensive experimentation, we showcase the heterogeneity and diversity of the simulation dataset. The evaluation results also revealed that all the data-driven models performed poorly in unseen scenarios, highlighting the urgent need to develop a robust and generalizable model for predicting crowd flow in indoor spaces.

In public spaces such as city centers, train stations, airports, shopping malls, and multi-modal hubs, accurately predicting pedestrian flow is crucial for effective crowd management e.g. congestion prevention and evacuation planning. Traditional microscopic simulation models offer fine-grained insights by simulating each pedestrian individually, but they are computationally intensive and typically used at the planning and design stage, making them unsuitable for real-time interventions in high-demand scenarios. Macroscopic models, on the other hand, reduce computational cost by aggregating pedestrian behavior and solving partial differential equations, but they typically require estimates of traffic states such as density and speed — quantities that are difficult to measure accurately in practice. Additionally, as the complexity of these physics-based models increases, their computational feasibility for real-time use becomes even more limited. Data-driven (machine learning) models provide a computationally efficient alternative, enhancing real-time prediction capabilities. However, they often require large historical datasets to generalize well, and their performance can degrade under out-of-distribution (OOD) conditions. Moreover, most black-box learning models lack interpretability and domain-specific insights, limiting their practical adoption. In this paper, we propose a novel pedestrian flow prediction model based on the theory of crowd diffusion. Our method estimates flow rates directly from sensor-observed data and infers both Origin–Destination (OD) demand and route choice probabilities to support real-time operations. To address the OOD challenge, we incorporate an online learning mechanism that continuously calibrates model parameters based on incoming observations.