Learning from transport data and testing intelligent systems.
Learning from a moving system
A road does not evolve in isolation. Speed on one link can be affected by traffic arriving from upstream, queues spilling back from downstream, and conditions on neighboring links. Forecasting several steps ahead therefore requires more than extending each road's own recent trend.
Watch film
Large language models can generate many plausible travel choices, but plausible is not the same as representative of real travelers. In the same choice situation, a human sample and an artificial sample may favor different options. If those differences are ignored, adding more synthetic responses can make a demand model more confident about the wrong behavior.
Watch film
Crashes are infrequent, yet road agencies must decide where to investigate or intervene before the next one occurs. That creates a prediction problem: can the conditions of a road segment help estimate its future crash count? Traffic volume, geometry, and other recorded features may interact in ways a simple straight-line rule misses.
Watch film
Traffic sensors do not provide a perfect record. Some readings vanish at random, and a broken detector may leave a whole stretch of time blank. Filling those gaps matters because a missing value can distort both the picture of past congestion and the models trained to forecast it.
Watch film
Conventional traffic modeling starts with assumptions about demand and road costs, then computes how travelers distribute themselves across a network. In practice, planners often face the reverse situation: the network and some traffic observations are known, but key behavioral or congestion parameters are not.
Watch film
Traffic speed has a memory. A slowdown observed now may be part of a brief fluctuation or the beginning of a longer queue, and those possibilities look different when recent history is taken into account. Short-term prediction is therefore a question of which past observations should remain influential.
Watch film
Ride requests appear unevenly across a city and change from one period to the next. A quiet zone can become busy while a neighboring zone empties, leaving drivers in the wrong places if dispatch relies only on the latest request count.
Watch film
An autonomous vehicle may drive for an enormous distance before encountering a truly dangerous situation. Testing only under ordinary traffic would make evidence about rare failures painfully slow to collect. Yet filling a test with threats creates another problem: failures observed in that artificial setting cannot be read as the natural crash rate.
Watch film
