A Machine Learning-Based ETA Estimator for Wi-Fi Transmissions

Recent advancements related to device to device (D2D) communication make it possible for a transmitting node to dynamically select the interface to be used for data transfers locally, without traversing any network infrastructure. In this scenario, a controller is identified, whose goal is to manage the D2D connection after its establishment. The software defined networking paradigm makes it possible to select this controller node via software: a device becomes the master node of a Wi-Fi-direct network, whereas the remaining units, i.e., the clients, can exchange data with other devices through the master. This paper develops a machine learning-based prediction algorithm for the aforementioned scenario, in which multiple elements, while receiving data from the controller, require an accurate on-the-fly estimation of the remaining transmission time, i.e., the estimated time of arrival. Different machine learning approaches are considered for this task, with the goal of exploiting only the information available at each client, without modifying any standard communication protocol. This information is critical when, for instance, a mobile user needs to decide whether or not to delay a data transfer, based on the load of the network and on the residual time under radio coverage from an access point.