- March 31st (Tuesday)
- Room 1509, National Institute of Informatics
- Dr. Philippe Owezarski
Research Director in LAAS-CNRS
- Wireless network monitoring: At the frontier of signal and data
- The multiplication of wireless networks may results in more medium contention and stronger spectral pollution. To maintain the communication quality, nodes will have to be more adaptive and take benefits of the available information about the state of the medium and of the communications. In that optic, this presentation will address 3 contributions that respectively tackle the following fields of wireless networks research: measurements, modelling and development of new solutions for nodes adaptability.
The first contribution concerns the conception and implementation of a measurement bench that complies with the requirements of these studies (i.e. it must allow cross-layer measurements starting from the PHY layer and offer maximal control over the environment and its perturbations).
Our solution is to perform the communications inside the RF protected environment of an anechoic room. To interfere with these communications, a modulated AWGN noise is injected inside the room with the help of a directive antenna. Inside the room, measurements are made using WIFI devices and RF equipments.
The second contribution aims to take advantage of the experimental platform measurements to improve the realism of ns-3 simulations. In fact, despite a recognized lack of realism, network simulators such as ns-3 are currently used to test new wireless protocols and applications.
In that optic, we propose a root cause analysis (RCA) model designed to detect configuration, implementation or modelling anomalies between the simulator and the experimental bench. The application of this method results in a major improvement of the realism of the ns-3 WIFI model.
The third and last contribution consists in the application of three machine learning algorithms, namely SVR, k-NN and DT, in order to predict the IP throughput measured on a wireless link. The estimations are based on SNR, RSS and noise measurements at the receiver side The different algorithms are evaluated according to the accuracy of their estimations but also according to their functional characteristics (e.g. model size, training time, ...). The results show that the SVR and DT algorithms used in conjunction of the SNR metric obtain the most accurate estimates. Moreover the SVR and DT algorithms are respectively the most efficient in terms of memory and time.
*Please replace [at] with @.