Abstract
The accuracy of a traffic simulation depends on the accuracy of the travel demandthat drives it, and this demand is not directly observable. Detectors record the vehicles
passing a point but not their origin or destination. This thesis addresses the inverse
problem of reconstructing a time-dependent origin-destination demand per route from
such aggregated observations. The conditions that matter most for planning are con
gested ones, and these are the hardest to reproduce, since congestion emerges from
vehicle interaction and is often caused by traffic on a road other than the one where
it appears. Because these quantities are emergent properties of individually simulated
vehicles, no analytical relationship between demand and observations exists. The task
is therefore formulated as derivative-free optimisation of a demand vector against a
weighted relative error, regularised by a penalty on abrupt temporal variation. TraffSim
was extended with a gRPC interface for external control, and a Python client encodes
candidate demands, drives the simulator and evaluates the objective. An entire opti
miser population is submitted as one batch of concurrently executed simulations, which
a sequential comparison showed to reduce elapsed time by close to the population size,
so a run finishes in one to two hours rather than a day. The problem was solved with the
covariance matrix adaptation evolution strategy. In ten optimisation runs on a small
network, the prescribed observations were reproduced to within one to three percent
wherever the targets were attainable. The most demanding prescribed a transient con
gestion episode on a road that must yield at an unmanaged crossing, so it could only be
produced by scheduling traffic on the crossing road. The method discovered this indirect
lever and reproduced the congestion to within 1.5% on average, but failed to reproduce
the subsequent recovery in every run, such congestion proving easier to create than to
clear. Reconstruction quality otherwise appeared to degrade as further segments were
constrained. The residual error proved to be concentrated on individual targets rather
than distributed, and on one segment the prescribed mean speed was never reached in
any of 4500 simulations across a wide range of flows. The apparent scaling behaviour was
therefore an artefact of one unattainable target rather than a property of the method.
Two conclusions follow. The reconstruction is expressive enough to produce a prescribed
congestion episode through an indirect physical mechanism. And since targets drawn
from real measurements are attainable by construction whereas hand-specified ones are
not, an infeasible target can dominate the achievable error while remaining invisible in
the aggregate figure, so synthetic targets should be checked for feasibility before use.
| Date of Award | 2026 |
|---|---|
| Original language | English (American) |
| Awarding Institution |
|
| Supervisor | Gerald Ostermayer (Supervisor) |
Studyprogram
- Mobile Computing
Cite this
- Standard