We consider the problem of using real-time floating car data to construct vehicle travel time prediction models meant to be used as input to routing algorithms for finding the fastest (time-shortest) path in the traffic network. More specifically we target the on-line car navigation systems. The travel time estimates for such a system need to be computed efficiently and provided for all short segments (links) of the roads network. We compare several fast real-time methods such as last observation, moving average and exponential smoothing, each combined with a historical traffic pattern model. Through a series of large-scale experiments on real-world data we show that the described approach yields promising results and conclude that specific prediction function form may be less important than a proper control of bias-variance trade-off (achieved by historical and real-time models combination). In addition, we consider two different settings for testing the prediction quality of the models. The first setting concerns measuring the prediction error on short road segments, while the second on longer paths through the traffic network. We show the quality and model parameters vary depending on the assessment method.
JavaScript jest wyłączony w Twojej przeglądarce internetowej. Włącz go, a następnie odśwież stronę, aby móc w pełni z niej korzystać.