A Comparative Study between Different Machine Learning Algorithms for Estimating the Vehicular Delay at Signalized Intersections

The delay at signalized intersections is a crucial parameter that determines the performance and level of service (LOS). The estimation models are commonly used to model delay; however, inaccurate predictions from these models can pose a significant limitation. Consequently, this study aimed to compare a wide array of machine learning algorithms, including Artificial Neural Networks (ANN), Random Forest (RF), decision tree, Support Vector Machine (SVM), K-Nearest Neighbor (K-NN), AdaBoost, Gradient Boost, XGBoost, and Partial Least Squares (PLS) regression. A comprehensive evaluation was undertaken across prediction accuracy, training-testing performance discrepancy, sensitivity to outliers, computational time cost, and model robustness. Additionally, the proposed methods were benchmarked against the Highway Capacity Manual (HCM), Webster, and Akçelik models. The results demonstrated that the RF model exhibited the most balanced performance across the specified criteria, with an average error below 4% and a rating of 35 out of 45 according to the proposed criteria. Moreover, the findings revealed that adopted analytical models should not be employed for vehicular delay estimation without calibration, as RMSE values were about 5 to 58 times higher than other models, varying by model.