The current trend of transforming old cities into smart cities has revealed many issues of the modern cities. One of the issues is the prevailing traffic jam on highways of the modern cities. The vehicular shock wave has been a problem on highways since it is one of the main causes of the traffic jam. The combination of heavy traffic and small traffic perturbations or unexpected driver actions are the main causes of shock waves. In order to alleviate road traffic caused by shock waves, it is crucial to have a system that predicts shock waves and informs them to the drivers. In this dissertation, we analyzed 6 months of freeway traffic data of Los Angeles, CA, provided by CalTrans PeMS (Performance Measurement System) and obtained the vehicular shock wave propagation speed of each freeway. Based on this information, we propose a machine learning approaches to predict shock waves. We utilize Hidden Markov Model (HMM) to predict if the shock wave will occur and propagate based on neighboring lanes' traffic information. Addtionally, HMM is used to estimate the probability of lane change from one lane to other lanes based on the occupancy of a lane. Baum-Welch algorithm is used to predict the parameters (occupancy and state).

We also utilized Deep Learning (DL) in order to predict the shock wave occurrence and propagation. We compared Stacked AutoEncoder (SAE), Deep Belief Networks (DBN), and HMM for the accuracy of the prediction of shock wave occurrences and propagation. These approaches have been tested on the same PeMS data sets and achieved good accuracy.

In the future, our models will be used to include modern collision prevention techniques (e.g., anti-shock wave strategies) to test their efficacy and help to reduce the number of potential accidents and save energy in the process. Also, our models can be used to improve traffic simulators to provide driving patterns that are close to real human's.