DeCoR: Design and Control Co-Optimization for Urban Streets Using Reinforcement Learning
The paper introduces DeCoR, a two-stage reinforcement learning framework that co-optimizes urban crosswalk layouts and adaptive signal control to significantly reduce pedestrian and vehicle delays on real-world corridors while generalizing to varying demands without retraining.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine a busy city street as a giant, chaotic dance floor. On one side, you have pedestrians trying to cross; on the other, cars trying to drive through. Currently, city planners often set up the dance floor (the crosswalks) based on old rules of thumb, like "put a crosswalk every 100 feet," and then hire a DJ (the traffic light) to play a fixed beat, regardless of how many people are actually dancing.
The paper introduces DeCoR, a new system that acts like a super-smart choreographer and DJ working together to redesign the dance floor while the music is playing.
Here is how it works, broken down into simple steps:
1. The Problem: The "Guesswork" Gap
Right now, we have great cameras and sensors that can count exactly how many people and cars are moving and where they want to go. But we rarely use that data to change the city layout. We measure the traffic, but we don't redesign the street to fit the traffic. This leads to dangerous situations where people jaywalk because the nearest crosswalk is too far away, or cars get stuck at lights that don't match the actual flow of people.
2. The Solution: A Two-Stage "Co-Optimization"
DeCoR uses a type of AI called Reinforcement Learning (think of it as a video game player that learns by trial and error) to solve two problems at once:
- Stage 1: The Architect (Design Agent). This AI looks at the map and asks, "Where should we put crosswalks, and how wide should they be?" It doesn't just guess; it uses a mathematical tool called a Gaussian Mixture Model. Imagine this as a "heat map" that learns the most popular paths people want to take (their "desire lines") and suggests placing crosswalks exactly where the crowd wants to go, rather than where a rulebook says they should be.
- Stage 2: The Conductor (Control Agent). Once the Architect suggests a new layout, the Conductor takes over. This AI controls the traffic lights. Instead of a fixed timer, it watches the real-time crowd and changes the lights instantly to let people and cars pass with the least amount of waiting.
3. The Training Ground: A Digital Twin
Before trying this on a real street, the team tested DeCoR on a 750-meter stretch of a university campus in the US. They fed the system real data from:
- Video cameras to count cars.
- Wi-Fi logs (from anonymous phones) to count pedestrians.
They ran thousands of simulations in a digital world (using a tool called SUMO) where the AI tried different crosswalk layouts and light timings. Every time the AI made a move that reduced waiting time for both people and cars, it got a "reward." Every time it caused a delay, it got a "penalty."
4. The Results: Smarter, Faster, Safer
The results were surprisingly effective:
- Fewer Crosswalks, Better Results: The real-world street had 7 crosswalks. DeCoR suggested removing some and keeping only 4, but placing them in smarter spots.
- Walking is Faster: Because the crosswalks were placed exactly where people naturally wanted to walk, pedestrians reached their nearest crossing 23% faster. They didn't have to walk out of their way as much.
- Waiting is Drastically Reduced:
- Pedestrians: Waited 79% less time at the lights compared to the old fixed-timer system. In fact, they waited almost as little as they would at an uncontrolled crossing (where you just walk across), but with the safety of a signal.
- Cars: Waited 65% less time. The lights adapted so cars didn't stop unnecessarily.
- The "Superpower" of Co-Optimization: The paper tested what happens if you change the street layout after the system is trained.
- If you train the DJ (Control) on a fixed floor plan, and then you suddenly add a new crosswalk, the DJ gets confused and traffic jams.
- But because DeCoR's DJ learned while the floor plan was changing, it became incredibly flexible. Even when the team added a new crosswalk that the AI hadn't seen before, the system kept traffic flowing smoothly, reducing wait times by 97% compared to a system trained the old-fashioned way.
The Bottom Line
DeCoR proves that if you listen to the data (sensing where people and cars actually go) and let the design (where the crosswalks are) and the control (the traffic lights) talk to each other, you can make cities safer and faster for everyone. It turns a rigid, rule-based system into a flexible, living organism that adapts to the crowd.
Note: The paper focuses strictly on improving travel time and reducing delays in a simulation based on real-world data. It does not claim to have eliminated accidents or fatalities in the real world yet, nor does it claim to work on every type of city street globally, but rather on the specific corridor they tested.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.