School of Science Department of Physics 74 Neural Information Processing for Highway Latency Prediction Supervisor: WONG Michael Kwok Yee / PHYS Student: HO Man Fung / PHYS-IRE SO Lap / PHYS-IRE JHA Shashwat / PHYS Course: UROP1100, Fall UROP1100, Fall UROP1100, Fall Using data extracted from the Taiwan Highway System, latency of highway journeys between different sources and destination points can be predicted. Significant factors causing the congestion were identified and machine learning models were trained to predict the latency. Previously, a classification method was introduced to preprocess the data into different clusters before training the machine learning model. Subsequently, it was found that by introducing the third dimension into the classification method, the accuracy in the latency prediction was improved. In this report, the effects of different choices of the third dimension on the latency prediction task were studied. Neural Information Processing for Highway Latency Prediction Supervisor: WONG Michael Kwok Yee / PHYS Student: HO Man Fung / PHYS-IRE Course: UROP2100, Spring From results collected from the research last year, it was found that the travel time (latency) between two detectors in Taiwan Highway System can be predicted via an analysis of traffic parameters collected by different detectors using machine learning models, including Extreme Gradient Boosting (XGB) and Gaussian Mixture Model. In this semester, the focus of the research is shifted to ways to increase the accuracy of the prediction by: a revision of inputs and outputs, modification of the training and testing sets and a refinement of the choices of the inputs to the Gaussian Mixture Model. Neural Information Processing for Highway Latency Prediction Supervisor: WONG Michael Kwok Yee / PHYS Student: SO Lap / PHYS-IRE Course: UROP2100, Spring In the task of latency prediction of highway journeys, we examined the error due to the deletion of inputs and studied the mutual information between inputs and outputs. We showed that the accumulation of vehicles in a highway segment and the past latency traversing the segment (and the speed of traversing) have high correlation with the current speed and latency. Compared with other choice of outputs, the mean latency is shown to be having the least mutual information with the inputs among mean, median speed, and latency. In addition, for the inputs to have significant mutual information with the outputs, it requires the segment length per sampling time to be lower than a threshold, which is likely to be related with the speed of vehicles traversing the highway segment.