Simulation-Driven Multi-objective Evolution for Traffic Light Optimization

Cacco, Alessandro; Iacca, Giovanni

doi:10.1007/978-3-030-43722-0_7

Cited by 1 publication

(1 citation statement)

References 9 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Yet, they typically require extensive training [1]. This downside can be mitigated by turning to evolutionary algorithms: versions specifically tailored to tackle complex problems related to urban transport have been successfully applied to automate traffic signal management [5,13], design bus route networks with a reduced environmental impact [8], locate electric vehicle charging stations [6], etc. Another promising approach to traffic prediction is related to transfer learning, a technique that enables transferring models trained on data collected from one area (source) to a neighbouring one (target), where traffic readings may not be available.…”

Section: Introductionmentioning

confidence: 99%

Predicting Normal and Anomalous Urban Traffic with Vectorial Genetic Programming and Transfer Learning

Hamilton¹,

Ekárt²,

Patelli³

2023

Applications of Evolutionary Computation

View full text Add to dashboard Cite

The robust and reliable prediction of urban traffic provides a pathway to reducing pollution, increasing road safety and minimising infrastructure costs. The data driven modelling of vehicle flow through major cities is an inherently complex task, given the intricate topology of real life road networks, the dynamic nature of urban traffic, often disrupted by construction work and large-scale social events, and the various failures of sensing equipment, leading to discontinuous and noisy readings. It thus becomes necessary to look beyond traditional optimisation approaches and consider evolutionary methods, such as Genetic Programming (GP). We investigate the quality of GP traffic models, under both normal and anomalous conditions (such as major sporting events), at two levels: spatial, where we enhance standard GP with Transfer Learning (TL) and diversity control in order to learn traffic patterns from areas neighbouring the one where a prediction is needed, and temporal. In the latter case, we propose two implementations of GP with TL: one that employs a lag operator to skip over a configurable number of anomalous traffic readings during training and one that leverages Vectorial GP, particularly its linear algebra operators, to smooth out the effect of anomalous data samples on model prediction quality. A thorough experimental investigation conducted on central Birmingham traffic readings collected before and during the 2022 Commonwealth Games demonstrates our models' usefulness in a variety of real-life scenarios.

show abstract