Machine Learning Models of Erroneous Driver Actions at Novel Interchange Configurations

Many state departments of transportation have designed and implemented grade-separated interchanges (GSI) to increase overall capacity at traffic intersections. As compared with traditional intersections, GSI designs are less familiar to drivers and, even with guidance signs, they are more likely to lead to erroneous driving actions. Based on the current literature, there is a need for predictive models of driver behavior based on GSI configurations and sign settings (i.e., use and placement of guidance signs). We evaluated multiple machine learning models for predicting whether an interchange design/configuration and sign settings are associated with erroneous driver actions, given different cognitive (workload and situation awareness) states. Results revealed a decision tree (DT) model to produce the highest prediction performance. Furthermore, based on importance scores for each factor in the DT, we found that interchange configuration was dominant in predicting erroneous driving outcomes, while cognitive workload and sign settings also played important roles in classification.