Non-parametric models with optimized training strategy for vehicles traffic flow prediction

Abstract

With the rapid development of information technology at the beginning of the 21st century, the traditional transportation system is rapidly transforming to Intelligent Transportation System (ITS). Meanwhile, as a powerful optimization tool for the applications’ performance under the framework of ITS, the traffic prediction model has gained much attention. The traffic prediction model belongs to the traditional time series prediction model, widely used in industry for business analysis, abnormalities detection, etc. There are two main categories of traffic prediction models — the parametric model and the non-parametric model. The non-parametric model develops rapidly in recent years due to the machine learning theory’s maturing and increment in computing power. Compared to the parametric model, the non-parametric model is more accurate and requires less advanced analysis of traffic patterns’ correlation, and can better handle a large amount of historical data. However, the non-parametric model also requires more powerful devices for training and implementation. The traffic prediction system’s implementation environment, built by on-road resources and off-road infrastructures, is relatively limited compared to the traditional data center. This paper uses several well-known state-of-the-art non-parametric models and their Deep Learning structures for traffic prediction, and evaluates the models’ performance on both freeway and urban-road dataset. Focusing on the model’s accuracy and training cost while deploying prediction models in a large-scale traffic network, this paper provides a novel optimized training strategy called CTS to reduce the implementation cost of a complex inner structure model. This approach could be further used to reduce the deployment cost, especially the training time, for the big data intelligent system using machine learning.