Wi-Fi throughput estimation and forecasting for vehicle-to-infrastructure communication

Abstract

Vehicles increasingly need to connect to external networking infrastructure, to support applications such as over-the-air updates, edge computing, and even autonomous driving. The ubiquity of IEEE 802.11 Wi-Fi makes it ideal for opportunistic vehicular access. However, that ubiquity also creates a problem of choice. In a heterogeneous Wi-Fi environment, where different networks coexist, it becomes important for vehicles to be able to pick the best-performing one. Focusing on delay-insensitive traffic, we equate network performance with throughput. To inform network selection we aim to first estimate current throughput, and then forecast its evolution through time. In order to avoid introducing load onto the network, we estimate throughput using only passively observable variables such as signal strength. We used Symbolic Regression (SR) and an Unscented Kalman Filter (UKF) to develop a computationally inexpensive estimation model — UKF-SR. We trained and tested this model using experimental data featuring 802.11n, ac, and ad networks. UKF-SR proved competitive against more expensive models such as shallow neural networks. To predict future throughput, we explored both general time-series forecasting models such as Autoregressive Integrated Moving Average (ARIMA), and domain-specific ones based on mobility information. The latter clusters historical throughput according to attributes such as vehicle position and direction of movement, using the cluster’s average as the forecast. An evaluation using experimental data showed the mobility-based models to meaningfully outperform general forecasting.