Abstract:
|
In this project, we were interested in using the vehicle speed obtained from a naturalistic driving dataset to develop a prediction model of whether a human driven vehicle would stop before executing a left turn. Such a model is a small step towards facilitating communication between driverless vehicles and human driven vehicles. Preliminary analyses showed that Bayesian additive regression trees (BART) produced less variable AUC values compared to Super Learner. BART also produced higher AUC values compared to linear logistic regression and non-linear logistic regression. Unfortunately, BART was designed for independent observations, but our dataset consists of longitudinal observations. Therefore, we extended BART to handle longitudinal prediction by adding a random intercept. We then used a simulation study to determine the prediction performance, bias, and 95% coverage of our proposed model for correlated binary outcomes. Lastly, we successfully implemented our proposed random intercept BART model to our dataset and found significant improvements in prediction performance compared to BART, linear logistic regression, and random intercept linear logistic regression.
|