Traffic

Problem

Long traffic jams are usually attributed to demand that exceeds a road's capacity for a long time. Traffic data show another mechanism: a short burst of demand can push a road segment past a critical density, after which the segment lets out fewer vehicles than arrive, and the jam outlasts the burst that caused it. If jams form this way, an operator has to keep each segment below the critical density, and has to do so without the new roadside hardware that most control methods need.

Approach and findings

We first measured how jams form, using loop-detector data from New York and speed data from Nairobi and São Paulo. Hourly data are too coarse to show a jam forming, so we derived each road's traffic curve from the distribution of its speeds, and found that 67% of the hourly slowdowns in New York contain a sudden jam in the minute-level data. Road segments in Nairobi spend 1.89 hours a day in jams on average, against 0.67 hours in New York and São Paulo.

Self-Regulating Cars acts on the critical density with an advisory speed sent to in-vehicle apps. A reinforcement-learning policy chooses one of three maximum speeds for each stretch of road, and it is penalised whenever a stretch exceeds the critical density. In simulation of a highway network in Mainz, Germany, the policy raised the number of completed trips by 5% compared with no control and lowered average delay by 7.7%, or by 12.6% for a version trained on a different traffic pattern. All three traffic-signal baselines had lower throughput than no control.

Two working papers take the idea further. The first removes the central controller: every equipped vehicle computes the same network-wide speed plan from a shared traffic feed, and an onboard camera corrects the advisory when the road ahead approaches the critical density. The system runs on about US$500 of commodity hardware, which we tested on public roads, and in simulation on the Mainz network it prevented the jam in 55 of the 74 runs that would otherwise have jammed. The second paper gives guarantees for decentralized admission control in merge networks with capacity drops, delayed actuation and partial compliance. It characterizes the largest set of states the controller can keep safe, and proves that, when a positive recovery margin exists, the same controller returns every road cell to its safe region within a bounded time.

Papers

Understanding Sudden Traffic Jams: From Emergence to Impact
Ankit Bhardwaj*, Shiva Iyer*, Sriram Ramesh, Jerome White, Lakshminarayanan Subramanian
Development Engineering, 2023. [Paper]

Self-Regulating Cars: Automating Traffic Control in Free Flow Road Networks
Ankit Bhardwaj, Rohail Asim, Sachin Kumar Chauhan, Yasir Zaki, Lakshminarayanan Subramanian
AAAI Conference on Artificial Intelligence (AAAI), AI for Social Impact track, 2026. Oral presentation. [Paper] [Code] [Video]

A Decentralized Speed Advisory System with Global Traffic Awareness and Local Visual Sensing
Ankit Bhardwaj, Denis Rybkin, Anirudh Sivaraman, Lakshminarayanan Subramanian
Under review.

Robust Admission Control for Cell Transmission Networks with Capacity Drops and Delays
Ankit Bhardwaj, Lakshminarayanan Subramanian
Under review.

Collaborators

• Rohail Asim (NYU)

• Sachin Kumar Chauhan (IIT Delhi)

• Shiva Iyer (NYU / Toyota ITC)

• Jerome White (World Bank)

• Yasir Zaki (NYU)

• Anirudh Sivaraman (NYU)

• Denis Rybkin

• Lakshminarayanan Subramanian (NYU, advisor)