Abstract
The open streets initiative "opens" streets to pedestrians and bicyclists by closing them to cars and trucks. The initiative, adopted by many cities across North America, increases community space in urban environments. But could open streets also make cities safer and less congested? We study this question by framing the choice of which streets to open as a reinforcement learning problem. In order to simulate the impact of opening streets, we first compare models for predicting vehicle collisions given network and temporal data. We find that a recurrent graph neural network, leveraging the graph structure and the short-term temporal dependence of the data, gives the best predictive performance. Then, with the ability to simulate collisions and traffic, we frame a reinforcement learning problem to find which streets to open. We compare the streets in the NYC Open Streets program to those proposed by a Q-learning algorithm. We find that the streets proposed by the Q-learning algorithm have reliably better outcomes, while streets in the program have similar outcomes to randomly selected streets. We present our work as a step toward principally choosing which streets to open for safer and less congested cities.
Introduction
"Open Streets" initiatives have gained significant popularity across North American urban centers as a means to reclaim community space. However, the impact of these initiatives on public safety and traffic flow remains a subject of intense debate. In cities like New York, street selection is typically driven by community proposals—a reactive process that can favor affluent neighborhoods and lacks a principled understanding of the "network effect." Closing a single street segment inevitably forces traffic onto neighboring roads, potentially shifting congestion and collision risks elsewhere. This research moves beyond political proposals toward a data-driven optimization strategy to identify which street closures actually improve urban safety and efficiency.
Methodology
The researchers developed a two-part framework to evaluate and optimize street closures in Manhattan, treating the problem as a reinforcement learning task.
Collision Prediction (The Simulator)
To accurately simulate the impact of a closure, the team first built a Recurrent Graph Neural Network (RGNN) to predict traffic collisions. The "Graph" component captures the spatial dependencies of Manhattan’s 19,391 road segments, while the "Recurrent" component accounts for temporal variables like weather and traffic patterns. This model serves as the environment for the reinforcement learning agent.
Street Selection (The Optimizer)
The choice of which streets to open was framed as a Deep Q-Learning (DQN) problem. An agent learns to select road segments that maximize a "reward" defined by the combined reduction in total collision risk and traffic density. To maintain realism, the simulation employs a $k$-weighted shortest-path algorithm to reroute traffic whenever a segment is closed to vehicles.
Data Sources
The study integrated several high-resolution datasets from New York City:
- NYC Open Data: Official police reports for motor vehicle collisions.
- Infrastructure Data: A detailed road-bed map of Manhattan.
- Taxi Trip Data: 10 to 15 million monthly trips (2013–2015) used to infer city-wide traffic flow.
- Weather Data: Hourly NOAA updates to account for environmental driving hazards.
Results and Analysis
Model Performance
The RGNN achieved a macro average recall of 0.76 for collision prediction, significantly outperforming state-of-the-art baselines such as DSTGCN and Graph WaveNet. This high predictive accuracy allowed for a robust simulation of how traffic changes affect safety.
Policy Evaluation
The most striking finding was the comparison between the current NYC Open Streets program and the RL-optimized approach. The Q-learning strategy consistently produced positive rewards, indicating a net reduction in both traffic and accidents. In contrast, the historical NYC Open Streets selections performed similarly to—and occasionally worse than—randomly selected streets.
Discussion and Insights
Beyond raw performance, the study revealed several critical insights for urban planners:
- Geographic Equity: RL-selected streets were more geographically diverse than the current program, which is heavily concentrated in Lower Manhattan. This suggests that algorithmic optimization can lead to a more equitable distribution of public space.
- Directional Bias: The model found that opening East-West streets often creates more disruption than benefit. Manhattan’s traffic is optimized for North-South travel; disrupting cross-streets frequently forces cars into less efficient routes.
- Cognitive Load Hypothesis: Counter-intuitively, the model associated higher car counts and slower travel times with decreased collision risk. The researchers hypothesize that drivers are more engaged and cautious in slow, heavy traffic, whereas fast-moving, sparse traffic may lead to decreased attention and higher-speed impacts.
Conclusion
This research demonstrates that Deep Reinforcement Learning can provide a principled foundation for urban planning decisions. While current proposal-based systems may fail to improve safety or congestion, an optimized approach can identify street closures that benefit the entire network. Future work will look to incorporate "global" rerouting (modeling drivers using navigation apps) and broader objectives like air quality and local economic impact.
Summary updated