On Training Traffic Predictors via Broad Learning Structures: A Benchmark Study

A fast architecture for real-time (i.e., minute-based) training of a traffic predictor is studied, based on the so-called broad learning system (BLS) paradigm. The study uses various datasets by California Department Transportation, and employs variety standard algorithms (LASSO regression, shallow deep neural networks, stacked autoencoders, convolutional, recurrent networks) comparison purposes: all are implemented in MATLAB same computing platform. demonstrates BLS process two-three orders magnitude faster (tens seconds against tens-hundreds thousands seconds), allowing unprecedented capabilities. Additional comparisons with extreme machine architecture, algorithm sharing some features BLS, confirm least-square as compared to gradient training.