Application of Online Regression Tree Induction to Forecast Traffic Flows

Abstract

Modelling efforts aiming at predicting traffic flows accurately should be capable of handling a continuous stream of data while being able to account for structural changes. To this end, the current research proposes three regression tree induction algorithms, including a batch, incremental and hybrid batch/incremental technique. These algorithms are discussed, implemented and evaluated based on their predictive power and coputational requirements. The incremental and batch/incremental algorithms prove to be particularly suited as on-line learning approaches, whereas the batch algorithm needs to be adapted in order to handle a continuous stream of data. The outcomes of the batch/incremental and batch algorithms ae comparable and these algorithms also prove to forecast traffic flows better than a selected baseline model. Additionally, the incremental, and batch/incremental approaches show to consume considerably less memory capacity and computational time compared to the adapted batch algorithm.