Multiobjective Selection of Input Sensors for SVR Applied to Road Traffic Prediction

road traffic forecasting, multiobjective feature selection, multiobjective genetic algorithms

Modern traffic sensors can measure various road traffic variables such as the traffic flow and average speed. However, some measurements can lead to incorrect data which cannot  further be used in subsequent processing tasks such as traffic prediction or intelligent control. In this paper, we propose a method selecting a subset of input sensors for a support vector regression (SVR) model which is used for traffic prediction. The method is based on a multimodal and multiobjective NSGA-II algorithm. The multiobjective approach allowed us to find a good trade off between the prediction error and the number of sensors in real-world situations when many traffic data measurements are unavailable.