A Federated Machine Learning Approach to Predicting Traffic Flow for Virtual Traffic Lights

In order to anticipate traffic, modern traffic systems now employ non-parametric techniques like machine learning. The traffic management system’s execution depends on sensory data sent to it, and a minor infrastructure breakdown, like a power outage, can significantly alter traffic patterns in the area. When this failure happens at a busy crossroads, it increases the time that cars idle, creates a traffic bottleneck, and may even result in a crash. A self-organizing traffic system called Virtual Traffic Light can be used to solve these problems: the vehicles at the intersection collaboratively manage the traffic flow using data from every vehicle present. Federated machine learning can be adopted when the vehicles collaborate to predict traffic, because data privacy is at the center of its operation. In this paper, I evaluate the ten-model zoo of such a system on traffic data from Austin, Texas, comparing centralized, federated, and local training at a single shared training budget and reporting every result against naive baselines on a held-out month. All ten models beat a persistence baseline, the best by 27.9%. The model that is best to deploy, however, is not the one that is most accurate centrally, and the usual explanation does not apply here: the two most accurate models are recurrent networks that can be federated, and at two clients they lose under 2% and remain the best available. The decisive variable is how many vehicles contribute. Across a sweep from 2 to 25 federated clients, the recurrent models lose 276% and 289% of their accuracy and fall below the naive baseline between four and five clients, while the two multi-layer perceptrons lose 14.5% and 18.6% and never fall below it. Only two federated models stay below the baseline under both partitionings at every client count: a multi-layer perceptron occupying 8 KB, 4.9% behind the best model under centralized training, and a 1 KB Ridge model, which is less accurate but almost untouched by federation. Federation is not the best way to train once each participant holds months of its own approach, though: a model trained on that history alone is more accurate than any federated model, at this intersection and at all nine others I tested. A federated model earns its place with a participant that has just joined, which it serves better than persistence for its first days. I further show that how clients are partitioned matters as much as which model is used, since splitting by approach, which is the situation a virtual traffic light is actually in, is considerably harder than splitting at random, and that federation’s cost is concentrated at peak demand: over the whole test month, it is 11.7% in the lower half of the demand distribution and 92.0% in the top decile. Finally, I tested the running system and not only the models. In eighteen live multi-node runs over three transports, some with packet loss, the nodes coordinated with no central server. There was one master at every sampled instant, every federated round finished, and every cycle took well under the 200 ms budget. None of the 1,413 validated plans led to a conflicting green on the road. This depends on the single master. With the single-master safeguards off, every plan still passed validation, yet conflicting greens showed on the road for 150.5 s in total across the same eighteen runs, and splitting the network on purpose caused them as well. A deployment must keep to one master and stop giving green when it cannot. In a queue model replaying the measured demand, adaptive control cleared 10.6% more vehicles than fixed timing at three times the measured demand and none more at the measured demand. The learned forecaster performed no better than naive persistence (within 0.2% at every demand level), since the two chose the same phase group in 97 to 98% of control steps.