A Comparison Between Adaptive Neural Networks Algorithms for Estimating Vehicle Travel Time

Abstract

The estimation of the time needed for a vehicle to reach a specific destination is one of the main focuses of navigation and Intelligent Transport Systems (ITS) as it helps both transit users and transit providers. Travel time estimation helps transportation providers to gain insight into evaluating travel routes, hence enhancing the transportation system reliability of their systems for transport users. In addition, travel time estimation helps in reducing the anxiety and stress for the travelers. Moreover, real time traffic data extremely impacts travel time estimation. Consequently, finding an accurate model for real time travel estimation is very crucial. Machine learning (ML) and its branch deep learning have proven to be efficient techniques to address this problem. Although there exists multiple ML models that estimate travel time, they are mainly offline models and they are fixed in size. Consequently, finding an adaptive online ML model is a vital task for real time travel estimation. This paper focuses comparing two adaptive online ML algorithms that operate in dynamic environment, namely multi-layer perceptron with hedge backpropagation and the greedy layer-wise pretraining. This paper shows that MLP with hedge backpropagation outperforms the greedy layer-wise pretraining algorithm. The mean square error percentages for MLP with hedge backpropagation and greedy layer-wise pretraining algorithm are reported to have values of 4.52% and 6.32%, respectively.