Traffic Conflict Prediction at Signal Cycle Level Using Bayesian Optimized Machine Learning Approaches

Abstract

This study develops non-parametric models to predict traffic conflicts at signalized intersections at the signal cycle level using machine learning approaches. Three different datasets were collected, one from Surrey, Canada, and the other two from Los Angeles and Georgia, U.S.A. From the datasets, traffic conflicts measured by modified time to collision and traffic parameters such as traffic volume, shockwave area, platoon ratio, and shockwave speed were extracted. Multilayer perceptron (MLP), support vector regression (SVR), and random forest (RF) models were developed based on the Surrey dataset, and the Bayesian optimization approach was adopted to optimize the model hyperparameters. The optimized models were applied to the Los Angeles and Georgia datasets to test their transferability, and they were also compared to a traditional safety performance function (SPF) developed using negative binominal regression. The results show that all the three Bayesian optimized machine learning models have high predictive accuracy and acceptable transferability, and the MLP model is a little better than the SVR and RF models. In addition, all three models outperform the traditional SPF with regard to predictive accuracy. The model sensitivity analysis also show that the traffic volume and shockwave area have positive effects on traffic conflicts, while the platoon ratio has negative effects.