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

Abstract

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.