Supervisors: Thomas Begin (LIP), Razvan Stanica (CITI) Location: LIP or CITI laboratory, Lyon
Periods: 5 months starting from January 2020 or later Salary: 550€ per month
Title: Using Artificial Intelligence to better configure WLANs
WLANs (Wireless Local Area Networks) have become part of our daily lives. They are offered at many different places and provide Internet access to many user devices and applications. WLANs are typically based on IEEE 802.11 standard (commercially known as WiFi). In order to meet the increasing needs of WLAN users, IEEE 802.11 has undergone several amendments, mostly aimed at strengthening its performance and security. In particular, MAC (Medium Access Control) and PHY (Physical) functions (specified by IEEE 802.11) have been enhanced. Indeed, transmission technologies, defining the PHY layer of IEEE 802.11, have significantly evolved over the years using e.g., wider channels, higher-order modulations, multiple-input multiple-output antennas (MIMO).
Maybe to a lesser extent, the MAC layer has also undergone some transformations with the possibility of using the Request to Send / Clear to Send mechanism (RTS/CTS), smaller mandatory waiting periods before transmissions, as well as frame aggregation and block acknowledgment in the latest amendments of IEEE 802.11. Network densification represents another means to cope with the growing demand of WLAN users. Typically, densification is carried out by increasing the number of Access Points (APs) that serve the users
Problem statement
Despite newer amendments of IEEE 802.11 and network densification, WLANs may be strained to keep up with the tremendous growth of demand. In particular, WLANs remain prone to performance and management issues such as unfairness and inefficiencies that may especially occur in dense networks. In this project, our goal is to address part of these issues by making fine adjustments to one key mechanism of IEEE 802.11: Rate Adaptation. Rate Adaptation (RA) is a mechanism that mostly belongs to the PHY layer. It allows APs and user devices to change their transmission rate with regard to the current quality of the radio channel. In a nutshell, the better the radio channel, the higher the transmission rate. Current approaches to choose the transmission rates are typically based on preset thresholds regarding the Frames Loss Rate or the received Signal-to- Interference-plus-Noise Ratio (SINR) [LAC04,BIA08,SLA12].
We believe that having preset values for the thresholds of RA is a conservative approach that is likely to lead to suboptimal performance for the WLANs and its users. As a matter of fact, it has been shown that certain unfair or inefficient WLAN situations (e.g., symmetrical and asymmetrical hidden node, flow in the middle) can be simply addressed by setting other values for some MAC parameters [AZI11,NAR12]. More generally, we believe that the performance of WLANs can be significantly improved through a fine and dynamic tuning of RA parameters. Unfortunately, as far as we know, there is no such thing as a general rule for how to set their values [NAR11].
Scientific approach
The goal of this internship is to develop an approach to dynamically select adequate values for the IEEE 802.11 parameters related to the RA mechanism to the WLAN context. The search for an adequate setting for the RA parameters is made complex due to the vast number of parameters (e.g., the used amendment of 802.11, the channel transmission rate, the number of competing nodes, the Frame Error Rate (FER), the offered load, and the transport protocol to name a few) that may affect a WLAN behavior. This high-dimensionality contributes to hinder the finding of general closed-form expressions.
In this internship, we propose to explore a new approach to determine an adequate setting of the RA
parameters using a data-driven approach based on techniques of Machine Learning (ML) in Artificial Intelligence (AI). Our approach consists of three stages. First, we will build a large dataset of measurements that will serve as the training set. This dataset should include the attained throughput of WLAN devices (output) as well as any WLAN parameters that may significantly affect these values (input) such as transmission rate at the PHY layer, number of neighbors, SNIR, frame loss rate, etc.
Second, we will use ML techniques to discover a function that fits the mapping between the dataset output and the inputs. Lastly, WLAN devices will embed and use this learned function to predict (approximately) what will be their attained throughput under various possible settings of their RA, and then select their best option. Overall, we expect to cast our problem as a Nonlinear Regression problem that we will address using Artificial Neural Networks that can easily handle problems of large dimensions.
Note that we will be able to compare the outcome of our approach with those of a real implementation using the actual RA code used by Intel wireless cards [INT18].
State of the Art
Over the last decades, remarkable progress has been achieved in the area of Artificial Intelligence (AI) and Machine Learning (ML) but they have not yet fully percolated through the networking community. The following quote is an excerpt from the website of the 2018 Workshop on AI in Networks (WAIN) jointly organized with IFIP WP Performance1: AI and ML are currently being exploited in almost every scientific fields. However, computer networks have still a limited development and deployment of these techniques.
Most existing works where ML techniques have been applied to computer networks deal with issues such as node localization for indoor wireless networks (e.g., [FIG12, ZOU15]), the spectrum utilization in cognitive radios [E3], or maybe to a lesser extent, the detection of intrusion (e.g., [KOL16]). As far we know little has been done in the context of adjusting the parameters of RA mechanism with the notable exception of [XXX].
[LAC04] M. Lacage, M.H. Manshaei, T. Turletti. (2004). IEEE 802.11 rate adaptation: a practical approach. ACM MSWIM.
[BIA08] S. Biaz, S. Wu. (2008). Rate adaptation algorithms for IEEE 802.11 networks: A survey and comparison.
IEEE ISCC.
[AZI11] A. Aziz, D. Starobinski, P. Thiran. (2011). Understanding and tackling the root causes of instability in wireless mesh networks. IEEE/ACM Transactions on Networking.
[NAR12] B. Nardelli, E. W. Knightly. (2012). Closed-form throughput expressions for CSMA networks with collisions and hidden terminals. In IEEE INFOCOM.
[NAR11] B. Nardelli, J. Lee, K. Lee, Y. Yi, S. Chong, E. W. Knightly, M. Chiang. (2011). Experimental evaluation of optimal CSMA. In IEEE INFOCOM.
[INT18] https://github.com/torvalds/linux/tree/master/drivers/net/wireless/intel/iwlwifi
[SLA12] T.M. Slåen. (2012). Classifying Rate Adaptation Algorithms in IEEE 802.11 b/g/n Wireless Networks (Master's thesis).
[FIG12] C. Figuera, J.L. Rojo-Álvarez, M. Wilby, I. Mora-Jiménez, A.J. Caamaño. (2012). Advanced support vector machines for 802.11 indoor location. Signal Processing.
[ZOU15 E21] H. Zou, X. Lu, H. Jiang, L. Xie. (2015). A fast and precise indoor localization algorithm based on an online sequential extreme learning machine. Sensors.
[KOL16] C. Kolias, G. Kambourakis, A. Stavrou, S. Gritzalis. (2016). Intrusion detection in 802.11 networks:
empirical evaluation of threats and a public dataset. IEEE Communications Surveys & Tutorials.
[KAR17] R. Karmakar, S. Chattopadhyay, S. Chakraborty (2017). IEEE 802.11 ac Link Adaptation Under Mobility. In IEEE Conference on Local Computer Networks (LCN).
1 https://performance2018.sciencesconf.org/page/wain
