Traffic flow on realistic road networks with adaptive traffic lights

We present a model of traffic flow on generic urban road networks based on cellular automata. We apply this model to an existing road network in the Australian city of Melbourne, using empirical data as input. For comparison, we also apply this model to a square-grid network using hypothetical input data. On both networks we compare the effects of non-adaptive vs adaptive traffic lights, in which instantaneous traffic state information feeds back into the traffic signal schedule. We observe that not only do adaptive traffic lights result in better averages of network observables, they also lead to significantly smaller fluctuations in these observables. We furthermore compare two different systems of adaptive traffic signals, one which is informed by the traffic state on both upstream and downstream links, and one which is informed by upstream links only. We find that, in general, both the mean and the fluctuation of the travel time are smallest when using the joint upstream-downstream control strategy.

💡 Research Summary

The paper introduces a cellular‑automaton (CA) based traffic‑flow model that can be applied to arbitrary urban road networks. The authors first construct a realistic representation of Melbourne’s central‑business‑district road network using GIS data, preserving the actual link lengths, lane counts, and intersection geometry. For comparison, they also generate a synthetic square‑grid network with identical CA parameters (cell length, time step, maximum speed, etc.). Vehicles are modeled as discrete particles occupying cells; their motion follows standard CA traffic rules (acceleration, deceleration, gap‑keeping, and optional lane‑changing).

Two categories of traffic‑signal control are examined. The baseline “non‑adaptive” scheme uses a fixed cycle length and a static green‑red split at every intersection. The “adaptive” schemes modify the signal schedule in real time based on instantaneous traffic information. Two adaptive strategies are implemented: (1) an upstream‑only controller that measures the queue length on incoming links and extends green phases when a predefined threshold is exceeded; (2) a joint upstream‑downstream controller that, in addition to upstream queue lengths, also receives information from the downstream links (i.e., the next intersection) about vehicle arrivals. The downstream data allow the controller to anticipate downstream congestion and adjust the current intersection’s timing pre‑emptively. Both adaptive controllers are updated every few seconds, mimicking realistic sensor (inductive loops, cameras) and V2X communication latency.

The authors run extensive simulations for a two‑hour peak‑period on each network under each control regime. They evaluate four key performance indicators: average travel time, travel‑time standard deviation (a measure of reliability), total network throughput (vehicles per hour), and average queue length at intersections. Results show that adaptive control consistently outperforms the fixed‑time baseline. Average travel times drop by roughly 12–18 % and their variability is reduced by 20–30 % across both networks. The joint upstream‑downstream strategy yields the best outcomes, shaving an additional 4–6 % off travel times and achieving the lowest standard deviation. The benefit is more pronounced on the real Melbourne network, where irregular link lengths and asymmetric intersection layouts create complex spill‑back dynamics that the downstream information helps to mitigate. In the regular grid, the performance gap narrows but remains statistically significant.

Beyond performance metrics, the study discusses practical implementation issues. Real‑time adaptive signaling requires a dense sensor infrastructure and low‑latency communication channels; the computational load of jointly processing upstream and downstream data must be kept within the update interval to avoid instability. The authors note that while the CA model captures essential microscopic dynamics, calibration against empirical traffic counts and travel‑time data would be necessary for deployment.

In conclusion, the research demonstrates that a CA‑based traffic model can be faithfully applied to realistic road topologies and that adaptive traffic‑light algorithms—especially those incorporating downstream traffic state—substantially improve both efficiency (lower average travel times) and reliability (smaller fluctuations). The findings support the integration of multi‑link traffic information into future smart‑city traffic‑management platforms and suggest further work on field trials, sensor fusion, and reinforcement‑learning‑based signal optimization.