A Succinct Shockwave Diagram Encoding Design to Enhance AI Efficiency and Assurance

Abstract: A shockwave diagram is a graphical representation of traffic state evolution over time and space, which contains critical traffic dynamic information essential for traffic monitoring and control. For example, shockwave diagrams could be used for estimating queue length and travel time, which can further improve traffic signal timing and ramp metering strategies. Recent advancements in deep learning (DL) have provided opportunities to leverage shockwave diagrams more effectively by learning underlying traffic patterns, particularly in detecting traffic congestion and issuing alerts for impending traffic events. However, the tabular format of shockwave diagrams is not inherently compatible with neural network processing. To address this issue, our study first developed a novel encoding method, inspired by the way shockwaves partition the temporal-spatial plane into disjoint polygons reminiscent of puzzle pieces. This puzzle-based encoding could capture detailed traffic state information, including flow, density, and spatial position, as well as the adjacency relationships among these traffic states. According to the puzzle-based encoding result, we also designed an interpretable RNN model by customizing trainable matrices to retain physical meaning of encoded features during recursive learning. To evaluate the data quality of the encoded shockwave diagrams, we next designed four new metrics from data accuracy, completeness, consistency, and uniqueness dimensions. Built upon these metrics, we analytically compare the puzzle-based encoding with the widely accepted pixel-based encoding. Using Seattle's highway loop detector and baseball game schedule dataset in 2011, we validated our encoding method by feeding its encoding result of shockwave diagrams in a DL model to predict public events. The experiment results demonstrated that the puzzle-based encoding could enhance AI efficiency and assurance.