Traffic Lights Considered Harmful
I've been thinking about traffic lights as a coordination problem, and I keep arriving at a strange conclusion: they might be a local optimum we've been stuck in for 150 years.
The average American driver spends something like two to three months of their life sitting at red lights. That's not internet hyperbole. INRIX data shows roughly 10% of trip time at signals, with an average delay of 18 seconds per light and only about 64% arrival-on-green. Combine that with an hour of daily driving over a 60-year driving life and you get thousands of hours of waiting.
This represents a massive coordination failure. Traffic lights optimize for worst-case peak loads while imposing inefficiencies during the roughly 70% of the time when intersections operate below capacity. The question I've been exploring is whether intelligent vehicle coordination can do better.
The answer, based on simulation work I've done, appears to be yes. But the path there is stranger than I expected.
The fundamental problem with traffic lights is architectural. They're a centralized solution to what is naturally a distributed coordination problem. Each vehicle only needs to not hit other vehicles. That's it. The light is a crude proxy for solving this, one that worked well when vehicles had no intelligence of their own.
But now vehicles can be intelligent. And the interesting question becomes: can we solve intersection coordination in its natural distributed form?
Formally, this is a multi-agent partially observable stochastic game. Each vehicle chooses actions to maximize its utility while maintaining safety constraints. The challenge is the curse of dimensionality: with n vehicles, the joint action space grows as |A|^n, which makes centralized optimization intractable for real-time applications.
But the problem has more structure than it first appears. Most vehicle pairs never interact. Conflicts are resolved within seconds. This sparsity makes the problem tractable.
The approach I've been working with combines neural Model Predictive Control with priority-based scheduling. Classical MPC formulates intersection coordination as constrained optimization:
Subject to system dynamics, state and control constraints, and collision avoidance. The problem is that centralized real-time MPC scales poorly. The coupled, nonconvex collision constraints often require sequential quadratic programming with per-iteration cost that grows roughly cubic in horizon times agents.
The key insight is that optimal MPC solutions likely lie on a low-dimensional manifold in the full control space. If this manifold can be learned, we can replace the expensive optimization with a single forward pass through a neural network. The state encoding aggregates ego state, K nearest neighbors, and temporal context. With K fixed by a cap, the input dimension avoids exponential growth with total agent count.
I bound the policy output to [-1,1] with Tanh and scale to acceleration, yielding [-3, +3] m/s². A Control Barrier Function safety filter then projects into the physically feasible set.
For collision avoidance, I use Closest Point of Approach analysis with rectangular vehicle models. For two vehicles with positions and velocities, you can compute the time of closest approach and minimum separation distance analytically. The CBF provides mathematical safety certificates by enforcing a discrete-time barrier constraint that guarantees the system stays safe.
Vehicles negotiate intersection access using priority-based reservations. Each vehicle estimates its arrival window based on acceleration bounds, and a scheduler grants access to the highest priority vehicle whose window doesn't conflict with existing reservations.
The results are interesting. In simulation, the system achieves roughly 35% higher average speeds through intersections and 60% reduction in waiting times compared to traffic lights. Zero collisions in 10,000 steps, which by the Rule of Three implies a 95% upper bound of approximately 3×10⁻⁴ collisions per step.
To be clear, that's nowhere near the safety level we'd need for deployment. Aviation demands something like 10⁻⁹ per crossing. Demonstrating that empirically with zero events would require on the order of 3×10⁹ crossings. This is the verification gap between simulation and reality, and I don't want to minimize it.
There are several findings that surprised me.
First, successful coordination may require higher computational costs initially to achieve lower costs asymptotically. This creates a deployment paradox. The transition period requires over-provisioning precisely when adoption rates and ROI are lowest. If neural MPC requires roughly 20× more compute than traditional control, early adopters face a substantial "coordination tax" with unclear benefits until network effects emerge. That probably requires 60-80% autonomous vehicle penetration.
Second, this approach creates a natural moat against commoditization. Unlike traditional traffic engineering, which can be copied, neural policies emerge from expensive simulation and cannot be easily reverse-engineered. This suggests a future where intersection management becomes a proprietary service rather than public infrastructure. I find that somewhat troubling, though I'm not sure what to do about it.
Third, most of the efficiency gains only matter in high-density urban cores. Most intersections operate below capacity most of the time. The complexity costs may not be justified except where traffic is genuinely heavy.
I should acknowledge the limits of what I've shown here. Simulation is not reality. Sensor failures, adversarial attacks, edge cases, weather, pedestrians, cyclists, legacy vehicles that don't coordinate: these are all problems I haven't addressed. The tail risks may dominate the safety analysis in ways that don't show up in clean simulations.
The most likely outcome is probably hybrid. Coordination-based management for new high-density developments, traditional signaling persisting in legacy infrastructure. The transition will be gradual and expensive.
But I do think the core insight is real: traffic lights are a centralized patch on a distributed problem, and as vehicle intelligence increases, the patch becomes increasingly awkward. Whether we actually make the transition depends on economics and politics as much as technology.
What stays with me is the sheer scale of the coordination failure we currently accept. Months of human life, exposed to the elements of waiting. It's one of those things that seems inevitable until you realize it isn't.