Research On Real-Time Traffic Information Forecast Algorithm Based On AOSVR

The successful implement of the intelligence transportation system (ITS) depend on forecasting real-time network transportation condition and analyzing network function of transportation planning layer and management layer by solid traffic analysis tools. Many subsystems of ITS, such as Advanced Transportation Management System (ATMS), dvanced Traveler information system(ATIS), Advanced Public Transportation System (APTS) and Emergency Management System (EMS), all lie on real-time and accurate forecast for current and future traffic information in a large-scale transportation network. Therefore, it is extremely necessary to do a deep research on traffic information forecast and build a real time traffic information forecast system to provide effective information for ITS. Much different from other forecast applications, the short-time traffic forecast highly demand real-time forecasting capability. It is required to finish many complicated works in a short time. Timeliness is of great importance.Statistical Learning Theory (SLT) is a kind of theory which works over theories of machine learning in the condition of small sample. Support Vector Machine (SVM), a novel technique of the statistical learning theory proposed by Vapnik in 1995, provides good solutions to carry out the problems of small sample, non-linearity, and so on. It has been generally applied in function approximation, pattern recognition and many other fields.Real-time forecast algorithms are studied in this paper. In order to improve time efficiency of forecast, a real-time forecast model with capability of on-line learning was proposed on the basis of AOSVR. And in the same while, a simplified computing method of Sigmoid kernel based on cloud model was developed.Through this kernel method, the calculation process is simplified, the speed of calculation is increased and the practical realization on hard ware is reinforced. The actual implementation has showed that the proposed model is superior to traditional SVMs and original AOSVR in both calculating rate and accuracy when it is applied in a real-time scheme. And the good robustness of cloud-model based Sigmoid kernel is also proved. Compared to other kernels, the cloud-model based Sigmoid kernel can obviously improve time efficiency of the forecast model at the expense of small loss in accuracy.