other models and accurately detect the cause of unnecessary active scanning.
The work in [51] proposes an SDN-based framework for AP selection in WLANs, considering the QoS level required by users’ flows. The authors of [52] study the user association issue in an SDN–architecture. They developed heuristic algorithms that lead to high performance assuming unsaturated heterogeneous Markovian analytical model.
The authors of [53] propose an admission control mechanism for VoIP calls in WLANs. A ML algorithm is used to predict the voice quality considering different parameters at the data link layer such as fraction of channel time used for video and normal traffic as well as estimated frame error rate for video and normal traffic.
The authors of [54] leverage on the SDN paradigm to develop an algorithm that achieves effective distribution of traffic load in WLANs. The authors try to optimally distribute network resources and improve the overall performance.
2.7.2.2 Interference Mitigation
Power control is a well-known approach used to mitigate interference in wireless networks. In SDN-based management and control of WLANs, the centralized CP can be used to implement the mechanisms for power control to minimize interference through coverage optimization of WLANs cells.
Wireless interference classification is a process of identifying the type of wireless emitters exist in the local RF environment [55]. This is important for enabling coexistence of wireless technologies that operate in the same frequency band. ML-based solutions are being developed to achieve this goal.
In [56], the authors propose a RL mechanism for interference mitigation in small cell networks. The algorithm represents the state of each AP as a binary variable that indicates whether the QoS requirement is violated. The action is a selection of power values from a set of power values. The reward is defined as the achieved rate. The algorithm iterates until a predefined level of QoS is met.
The work in [57] develops a solution that uses ML-SVM for interference classification in wireless sensor networks from IEEE 802.11 signals and microwave ovens. Another deep learning approach for classification of WiFi, Zigbee, and Bluetooth was proposed in [58]. The authors defined fifteen classification tasks assuming a flat fading channel with additive white Gaussian noise. The research work of [59] compares different types of ML models for classifying signals, including deep feed-forward networks, deep convolutional networks, SVM and a multi-stage training algorithm.
In [60], the authors propose a ML-based framework for mitigating the effect of jammers in WLANs, called “DeepWifi”. The system consists of an RF front end processing unit which applies a deep learning-based auto-encoder to extract spectrum-representative features. The system leverages the advances in ML algorithms to enhance the performance and security in WLANs. A deep neural network is then trained to classify signals as idle, WiFi, or jammer. In standard WiFi, the user backs off backs off regardless of the type of interference. However, DeepWiFi which is able to classify signals backs off when the interference is from another WLAN user, allowing user to operate in degraded mode and still receive non-zero throughput.
2.7.2.3 Channel Allocation and Channel Bonding
Even with centralized control, optimal channel allocation problem in WLANs is difficult to be solved in an acceptable complexity level. Recently, researchers have been trying to leverage ML methods to find solutions in feasible time.
In [61], the authors propose a ML method for assigning channels to WLANs APs. The method is based on passive monitoring of data in each cell. Using ML, it calculates the performance loss due to interfering users and finds the best channels for the cells that minimize interference. The algorithm minimizes airtime usage of interfering links in neighboring cells. Due to the dynamic nature of WLANs, the process is repeated iteratively. The authors of [62] concluded that a central control of APs is needed even if the network is influenced by neighboring unmanaged APs. Their approach results in a self-organizing system for channel allocation in WLANs based on cooperation between APs. The authors show that the proposed system leads to a stable network of high performance.
In [63], the authors use ML techniques to learn implicit performance models from real-world measurements. The techniques do not need to know the details of interacting parameters. The authors used the developed model for channel allocation and power control.
2.7.2.4 Latency Estimation and Frame Length Selection
Latency is a key factor that impacts the performance of modern mobile applications. In [64], the authors found that, latency depends on three main parameters: Channel utilization, the number of online devices, and the SNR. WiFi latency can be modeled using these related factors. The authors developed and compared the performance of supervised ML-based algorithms used to measure, characterize, and predict delay in large-scale WLANs. Training is implemented using data sets obtained from field measurements.
Selecting a proper frame length is an important issue in WLANs, where it impacts the performance and users’ QoE in the network. The selection problem requires advanced techniques able to utilize information on practical settings in real-time.
The work in [65] proposes an SDN-based solution for frame length selection in WLANs. The system proposes inclusion of ML techniques in SD-WLANs to optimize the selection of frame length for each user based on channel conditions as well as overall performance indicators. The supervised learning approach is used, where the algorithm is deployed on the management plane of the SDN architecture. The CP periodically feeds the algorithm with network knowledge about channel conditions and users’ state.
The research work of [66] proposes a ML-based approach for the implementation of QoS management model in wireless networks. The ML system uses both supervised and unsupervised algorithms to identify key quality indicators for network users which represent an estimation of the quality as perceived by users considering influencing factors. Also, the ML concept is used for providing information about areas where corrective actions are required.
2.7.2.5 Handover
Transparent handover with minimum overhead is still an open issue in WLANs. Though the 802.11r standard developed protocols that help implementing seamless handover between WLAN cells, still APs and users need to be highly engaged in the handover process. This impacts the performance of APs, especially in dense deployments; wherein handover rate is expected to be high. With the advancement in AI and ML; coupled with the evolving SDN technology, researchers are trying to develop methods that allow low cost and successful transparent mobility among WLAN cells.
An example effort is published in [67], where the authors developed an SDN-based solution for controlling and managing handover in WLANs. The proposed solution allows the devices to seamlessly move across cells without losing the QoS level.
The researchers in [68] developed a framework to optimize the handover process and balance the network throughput and handover rate. Unsupervised ML algorithm is used to classify users according to their mobility patterns. Then, deep RL is used to optimize the handover process in each cluster. The received signal power by the user from APs is used as the state vector. The reward is considered to be the weighted sum between the handover rate and the throughput.
In [67], the authors developed and tested an SDN-based solution for providing seamless handover in WLANs based on virtual APs. The solution maintains QoS requirements of real time applications in terms of packet loss and delay.
2.7.3
