Topology-regularized universal vector autoregression for traffic forecasting in large urban areas.

Autonomous vehicles are soon to become ubiquitous in large urban areas, encompassing cities, suburbs and vast highway networks. In turn, this will bring new challenges to the existing traffic management expert systems. Concurrently, urban development is causing growth, thus changing the network structures. As such, a new generation of adaptive algorithms are needed, ones that learn in real-time, capture the multivariate nonlinear spatio-temporal dependencies and are easily adaptable to new data (e.g. weather or crowdsourced data) and changes in network structure, without having to retrain and/or redeploy the entire system. We propose learning Topology-Regularized Universal Vector Autoregression (TRU-VAR) and examplify deployment with of state-of-the-art function approximators. Our expert system produces reliable forecasts in large urban areas and is best described as scalable, versatile and accurate. By introducing constraints via a topology-designed adjacency matrix (TDAM), we simultaneously reduce computational complexity while improving accuracy by capturing the non-linear spatio-temporal dependencies between timeseries. The strength of our method also resides in its redundancy through modularity and adaptability via the TDAM, which can be altered even while the system is deployed. The large-scale network-wide empirical evaluations on two qualitatively and quantitatively different datasets show that our method scales well and can be trained efficiently with low generalization error. We also provide a broad review of the literature and illustrate the complex dependencies at intersections and discuss the issues of data broadcasted by road network sensors. The lowest prediction error was observed for TRU-VAR, which outperforms ARIMA in all cases and the equivalent univariate predictors in almost all cases for both datasets. We conclude that forecasting accuracy is heavily influenced by the TDAM, which should be tailored specifically for each dataset and network type. Further improvements are possible based on including additional data in the model, such as readings from different metrics. [ABSTRACT FROM AUTHOR]