I Open at the Close: A Deep Reinforcement Learning Evaluation of Open Streets Initiatives

R. Teal Witter, Lucas Rosenblatt

Introduction

Researchers at NYU have developed a Deep Reinforcement Learning (DRL) framework to optimize "Open Streets" initiatives—urban programs that close roads to cars to create pedestrian spaces. While the social benefits of these programs are clear, their impact on city-wide logistics has remained a "black box." This paper addresses the lack of a systematic, predictive framework for evaluating how these closures affect the complex, interconnected web of urban traffic and public safety.

Currently, cities like New York choose which streets to open based on a manual proposal process, which can be influenced by local politics or neighborhood advocacy rather than data-driven safety metrics. This research investigates whether closing a street simply pushes traffic and accidents to the next block, or if it can actually make a city safer and more fluid.

Methodology

The authors split their approach into two distinct phases: collision prediction and street selection.

  1. Collision Prediction (The Simulator): To understand the "risk" of a street, they built a Recurrent Graph Neural Network (RGNN). The road network is treated as a graph where nodes are road segments and edges are intersections. Recurrent layers capture short-term temporal dependencies like weather and time of day. Unlike prior models that down-sample data to balance rare events, this model uses the entire dataset and weights the loss function to handle the extreme sparsity of collisions.
  2. Street Selection (The Agent): With a working simulator, the choice of which streets to open is framed as a Reinforcement Learning problem using a Deep Q-Learning (DQN) agent. When a road is "opened" (closed to cars), the system simulates traffic rerouting using $k$-weighted shortest paths. The agent is rewarded for reducing the sum of predicted collision risk and car density across the entire network.
Overview of the two-part framework: collision prediction via RGNN and street selection via Deep Q-Learning.

Datasets

The study focuses on Manhattan, NYC, using a massive multi-modal dataset from 2013–2015:

  • NYC Open Data: Detailed police reports of every motor vehicle collision.
  • NYC Department of Planning (LION): A comprehensive road-bed representation containing ~19,000 road segments and their attributes (speed limits, lanes, etc.).
  • NOAA: Hourly weather data from Central Park.
  • NYC Taxi & Limousine Commission (TLC): Data from tens of millions of taxi trips used to accurately infer traffic flow via start/end coordinates.

Key Insights and Ideas

  • The "Cognitive Load" Hypothesis: Interestingly, the model found that a higher "Number of Cars" and longer "Travel Times" were associated with decreased collision risk. The authors hypothesize that driving is cognitively easier in slower, congested conditions, whereas high-speed, low-traffic roads encourage riskier maneuvers.
  • Algorithmic Equity: The current NYC Open Streets program is heavily concentrated in wealthy areas. The RL agent proposed a much more geographically diverse distribution, suggesting that data-driven approaches could lead to more equitable urban planning.
  • The East-West Constraint: The model consistently gave low "value" to opening East-West streets. Because Manhattan’s grid is optimized for North-South flow, closing the limited East-West connectors causes disproportionate traffic chaos.

Key Results

The RGNN achieved a Macro Average Recall of 0.76, outperforming standard baselines like XGBoost (0.74) and specialized spatio-temporal models like DSTGCN (0.57). Crucially, the streets selected by the Q-learning agent provided consistently positive rewards, indicating a reduction in traffic and accidents.

Comparison of improvements in traffic congestion and safety across different selection methods.

Perhaps most strikingly, the streets currently in the NYC Open Streets program performed no better than randomly selected streets. This suggests that the current proposal-based system is missing significant opportunities to improve city safety.

Limitations and Future Work

The traffic inference relies on yellow taxi data, which may not perfectly capture the behavior of private vehicles or modern ride-share apps. Additionally, the simulation uses "local" rerouting for computational feasibility, whereas real-world drivers might adjust their entire route. Future work could integrate other vital urban metrics, such as air quality, noise levels, or economic impact on local businesses.

Summary updated

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 open streets initiative 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 already selected by the open streets initiative 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.
BibTeX
@inproceedings{witter2024i,
  author = {R. Teal Witter and Lucas Rosenblatt},
  title = {{I Open at the Close: A Deep Reinforcement Learning Evaluation of Open Streets Initiatives}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2024},
  doi = {10.1609/aaai.v38i20.30250},
}
Figure