Machine Learning Based Prediction with Parameters Tuning of Multi-Label Real Road Vehicles Characteristics

Abstract

The real-time traffic characteristics on the road network highly affect the safety conditions and the driving behaviors there. Early detection of crowded areas or hazardous conditions on the road network should affect the drivers' decisions and behavior to guarantee smooth and comfortable trips. Machine learning mechanisms have been mainly used for general prediction after extensive training processes. Over the road networks, trained machines could be really helpful to obtain instant predictions that assist drivers and autonomous vehicles there. However, the quality and efficiency of these machines are affected by several criteria including the quality of the used dataset and the tuning of the parameters of the regression algorithm. In this work, we investigate the performance of the most popular regression algorithms in terms of temporal prediction of the traffic characteristics in a real road scenario. Moreover, we optimize the regression algorithm by tuning the parameters using the grid search technique. From the experimental results, we can clearly notice the enhancements in predicting the traffic characteristics for different periods of time. We have observed that the number of neighbors, the distance, and the metric parameters' values are best tuned with the values of 4, 'Manhattan', and 'Distance', respectively, for the K-Nearest Neighbor (KNN) regression algorithm.