Random Vector Functional Link Networks for Road Traffic Forecasting: Performance Comparison and Stability Analysis

Nowadays, Machine Learning algorithms enjoy a great momentum in multiple engineering and scientific fields. In the context of road traffic forecasting, the number of contributions resorting to these modeling techniques is increasing steadily over the last decade, in particular those based on deep neural networks. In parallel, randomization based neural networks have progressively garnered the interest of the community due to their learning efficiency and competitive predictive performance. Although these two properties are often sought for practical traffic forecasting solutions, randomization based neural networks have so far been scarcely investigated for this domain. In particular, the instability of these models due to the randomization of part of their parameters is often a deciding factor for discarding them in favor of other modeling choices. This research work sheds light on this matter by elaborating on the suitability of Random Vector Functional Link (RVFL) for road traffic forecasting. On one hand, multiple RVFL variants (single-layer RVFL, deep RVFL and ensemble deep RVFL) are compared to other Machine Learning algorithms over an extensive experimental setup, which comprises traffic data collected at diverse geographical locations that differ in the context and nature of the collected traffic measurements. On the other hand, the stability of RVFL models is analyzed towards providing insights about the compromise between model complexity and performance. The results obtained by the distinct RVFL approaches are found to be similar than those elicited by other data driven methods, yet requiring a much lower number of trainable parameters and thereby, drastically shorter training times and computational effort.