Memory Sketching for Data-driven Prediction of Dynamical Systems

In this paper, we introduce a sketching method for data-driven prediction problems in an online setting. We show that sketching the memory of dynamical systems in a randomized way can achieve a constant time and space complexity in each update. We demonstrate the effectiveness of the proposed sketch-based data-driven prediction on trajectory prediction in vehicular traffic. We show that the sketching method can achieve high prediction accuracy with limited memory space, thus enabling online deployment.