Back to the Future: Predicting Traffic Shockwave Formation and Propagation Using a Convolutional Encoder-Decoder Network Mohammadreza Khajeh-Hosseini Department of Civil Engineering Texas A&M University College Station, USA mreza@tamu.edu Alireza Talebpour* Department of Civil Engineering Texas A&M University College Station, USA atalebpour@tamu.edu Abstract—This study proposes a deep learning methodology to predict the propagation of traffic shockwaves. The input to the deep neural network is time-space diagram of the study segment, and the output of the network is the predicted (future) propagation of the shockwave on the study segment in the form of time-space diagram. The main feature of the proposed methodology is the ability to extract the features embedded in the time-space diagram to predict the propagation of traffic shockwaves. Index Terms—Traffic Shockwave Propagation, Convolutional Encoder-Decoder, Traffic Prediction I. I NTRODUCTION AND BACKGROUND The boundary between two different traffic states is known as traffic shockwave. The driving dynamics change from one state to another as the speed of the vehicles and their spacing changes. The differences between the two traffic states can be mild such as a high-speed traffic stream reaching a traffic stream with moderate speed and density, or it can be significant when reaching a high density and low-speed traffic stream (e.g., congested area). In general, when the traffic state changes, the vehicles need to respond by adjust- ing their speed and acceleration. Currently, the approaches adopted for guidance (e.g., lane changing) of the autonomous vehicles involves the consideration of the current state of the surrounding vehicles in terms of their location and speed with limited attention to the response of the other vehicles to their surrounding environment and how the traffic state could evolve (e.g. formation of shockwaves). As a result, predicting the propagation of traffic shockwave in time and space can help in improving both the safety and performance of the autonomous vehicles. Considering the valuable information that the connected vehicles could provide, we are proposing a methodology to predict the propagation of traffic shockwaves that accounts for both the individual behaviors and the collec- tive change in the state of the traffic. The state of traffic is characterized by density, flow, and speed that change over time and space (i.e., along the roadway). Lint and Hinsbergen [1] classify traffic prediction methodologies into three general groups of naive, parametric and non-parametric approaches. Naive approaches apply to the methodologies that have no model assumption or parameters driven from data. The naive approaches either assume the future state remains similar to the current state or stays near the average of historical observations for the time of the prediction [2]. Parametric approaches refer to the methodologies that use a traffic flow model with parameters calibrated on historical data or in combination with new observations. The fundamental diagram in combination with first and second order traffic flow models are among the well studied parametric traffic flow models relating the macroscopic characteristic of traffic state [3], [4]. One of the challenges in the use of parametric models is the trade-off between accuracy and the complexity of traffic prediction models, especially when addressing the time-variant and abnormalities in traffic dynamics. Non-parametric approaches, on the other hand, are usually based on simple data-driven methodologies that do not explic- itly rely on a traffic flow model. There is a wide range of data analysis and machine learning approaches used in the field of traffic prediction. Some of the common approaches include linear regression [5], different classes of neural network [6], support vector regression [7], and time-series forecasting [8]. With the increase in the availability of data resources, the non- parametric approaches have gained more attention in the past decade. Most of the existing non-parametric approaches rely on aggregated and macroscopic traffic data such as flow, den- sity, and speed with limited studies focused on the trajectory of vehicles [9], [10]. With the increase in flow and density, the impact of unexpected driving behaviors on the state of traffic stream increases and more complex dynamics can be observed in the traffic flow. Consequently, capturing the interactions among vehicles can lead to a better prediction of traffic and change of states. There is a broad body of literature on traffic state prediction as well as comprehensive paper reviews such as [11] that are recommended for further study to interested readers. The connected vehicles’ technology provides the opportu- nity to disseminate useful data to share with the drivers, or arXiv:1905.02197v1 [cs.LG] 4 May 2019