Self-Learning Traffic Prediction in Los Angeles: A Continuous Learning Experiment

Some projects are really a question you cannot put down, wearing the costume of a software demo. This was one of them.

A while before I founded Wakeline, I spent my evenings on a question that still occupies me today. What would an intelligence look like if it never stopped learning? Not a model trained once and then frozen, but a system that keeps updating itself while it runs, from whatever the world happens to be doing right now. Traffic turned out to be a wonderful place to test the idea, and Los Angeles offered some of the most punishing traffic I could find.

The Experiment

I pointed my self-learning traffic software at a small piece of Los Angeles, the interchange where the Santa Monica and San Diego Freeways meet, and let it watch. Then I recorded an ordinary morning commute.

It is oddly satisfying to watch. The forecasts start out marked in blue. As the system reads the live conditions, the streets shift from green to red while congestion builds, and the predictions track the real flow road by road rather than as a single citywide average. When the commute winds down, the system simply stops forecasting. Nobody told it the rush hour was over. It worked that out from the signal.

Why It Was Not an Ordinary Traffic Model

The interesting part is not the prediction. Plenty of systems predict traffic. The interesting part is how it learned.

There was no pre-processed training set and no labelling exercise. I did not hand it a tidy historical dataset and wait for a training run to finish. It learned directly from the live stream of conditions, unsupervised, building its understanding of each road as it observed it. This is what people in the field now call online learning, learning from a stream one observation at a time, and it sits at the heart of what I would later spend my career on.

When the patterns shifted, it adapted in place. No retraining cycle. No redeployment on my end. It committed to a prediction only when it was confident enough to stand behind it, and it grew more capable the longer it watched. The behaviour you see in the video, the forecast quietly correcting itself toward reality, is the system updating its own understanding in real time rather than replaying something it memorised months earlier.

It also ran on a famously modest cloud setup. Because the system updates incrementally instead of grinding through enormous historical datasets, it costs very little to operate. That efficiency is the whole reason I could offer the Wolfsburg version to the public for free.

Why I Still Think About This

Looking back, the traffic work was a toy proving ground for an idea I now take very seriously. A live environment that never holds still. An intelligence that has to stay aligned with it without stopping to retrain. A cost structure that stays low because learning happens a little at a time rather than in one enormous batch.

That is precisely the bet behind the company I run today. At Wakeline we build continuous learning systems for environments far less forgiving than a freeway at rush hour. Our first product forecasts day-ahead electricity prices, where the signal drifts constantly and being wrong is measured in real money rather than a few minutes of commute. The freeway taught me the shape of the problem. The grid is where the architecture has to earn its keep.

There is a name for this now, continual learning, the ability of a system to keep learning after it is deployed without overwriting what it already knows. Most of the AI you use every day cannot do it. The model in your phone learned everything it will ever know before it met you, and it has not learned anything since. The little traffic predictor watching a Los Angeles interchange was, in its small way, an argument that things could work differently.

The questions I was playing with over those Los Angeles evenings turned out to be the right ones. I simply had not realised yet how far they would take me.