Simple Explanation (ELI5)

Picture a blind orchestra conductor trying to sync musicians by staring at the whole room: it sounds awful. Now imagine each musician listens only to their neighbor and adjusts their own tempo: a perfect melody emerges. The new traffic lights work exactly like that: each one talks only to the next one and knows how many cars will switch lanes, so green arrives exactly when needed.

Key Advantages & Current Limits

🟢 Pros: Shorter queues (-35% stops) and lightning-fast computation (<100 ms) thanks to distributed control, no more central bottleneck. 🔴 Cons: Requires detectors and sensors on every lane to feed the LSTM neural network, and the system has only been tested in simulation, not yet on real roads.

Future Outlook & Takeaways

Within 3-5 years we'll see the first pilot districts where traffic lights talk to each other while also managing pedestrians and public transport, integrating with connected cars to suggest the ideal speed and cut stop-and-go. For citizens it means shorter commutes, cleaner air, and less stress behind the wheel.

Traffic lights that 'talk' to each other: the math that makes...

Q: Future Outlook & Takeaways

Within 3-5 years we'll see the first pilot districts where traffic lights talk to each other while also managing pedestrians and public transport, integrating with connected cars to suggest the ideal speed and cut stop-and-go. For citizens it means shorter commutes, cleaner air, and less stress behind the wheel.

Article

100%

How many times have you sat at a red light while the lane next to you was empty? It's the classic urban traffic paradox: traditional traffic lights look at the road as a whole, ignoring that cars distribute differently between the overtaking lane and the regular lane. A fresh study on arXiv changes the game. Researchers built a "two-lane" model that simulates vehicle flow with unprecedented granularity: each lane has its own density, speed, and life.

On this base, they grafted a distributed Model Predictive Control (MPC) solved with the ADMM algorithm: in practice, each intersection calculates its own optimal green times by exchanging only essential info with neighbors, without a central brain that would slow everything down. The cherry on top is a data-driven module that predicts in real-time how many drivers will change lanes ("split ratios"), making the system as reactive as an expert driver.

Tested in VISSIM simulation on a real six-intersection network in Dalian, China, the system beat classic strategies for both traffic flow and computation speed. Result: shorter queues, less waiting, and an architecture ready for real-world deployment. The smart city is no longer a promise; it's becoming code running on controllers at the roadside.

Bonus

100%

The core innovation lies in spatial discretization: the classic LWR (Lighthill-Whitham-Richards) model treats the road as a single pipe. Here, authors extend the framework to two coupled pipes via source terms modeling lane-changing. The continuity equation for each lane i becomes ∂ρ_i/∂t + ∂q_i(ρ_i)/∂x = s_i, where s_i represents the net flow of vehicles entering/exiting lane i.

The distributed MPC control minimizes a local cost function (weighted sum of queues) subject to capacity and min/max green time constraints. ADMM (Alternating Direction Method of Multipliers) decomposes the global problem into per-intersection subproblems, coordinated by dual variables associated with boundary flows. Convergence is guaranteed by convexity of the relaxed problem.

The estimation of split ratios λ_i(t) (fraction of flow turning right/left/through per lane) uses a data-driven approach: an LSTM network trained on historical loop-detector data predicts λ_i(t+k|t) for the MPC prediction horizon. This avoids the stationary assumption typical of SCATS/SCOOT controls. VISSIM validation on a real Dalian network (6 intersections, 12 approaches, 24 lanes) shows an 18% reduction in average travel time and 35% fewer stops compared to offline-optimized Webster, with computation time per step < 100 ms on edge hardware, compatible with 1-5 second control cycles.