MACHINE LEARNING FOR PERFORMANCE PREDICTION OF CHANNEL BONDING IN NEXT-GENERATION IEEE 802.11 WLANS

Next-generation IEEE 802.11 WLANs are called to face the challenge of providing high performance under complex situations, e.g., to provide high throughput in massively crowded deployments where multiple devices coexist within the same area. To fulfill the strict requirements derived from novel use cases, features such as Multiple-Input Multiple-Output (MIMO), Spatial reuse (SR), or multi-Access Point (AP) coordination are being developed and incorporated into the newest amendments, namely IEEE 802.11ax (11ax) and IEEE 802.11be (11be) [5, 6].

One of the features that are receiving more attention is Channel Bonding (CB) [7, 8], whereby multiple frequency channels can be bonded with the aim of increasing the bandwidth of a given transmission, thus potentially improving the throughput. Since its introduction to the 802.11n amendment, where up to two basic channels of 20 MHz could be bonded to form a single one of 40 MHz, the specification on CB has evolved and currently allows for channel widths of 160 MHz (11ac/11ax). Moreover, CB is expected to support up to 320 MHz channels in the 11be amendment.

Figure 1 shows a snapshot of the 5 GHz band in Wi-Fi (particularly, U-NII-1 and U-NII-2 bands are shown), where basic 20 MHz channels can be bonded to form wider channels of up to 160 MHz. As it can be appreciated, the number of combinations for bonding channels is high, even for a small portion of the available spectrum¹¹1In the 5 GHz and 6 GHz bands, there are six and fourteen non-overlapping channels of 80 MHz, respectively. and under the constraint that only contiguous channels can be bonded. Moreover, novel bonding techniques combine Orthogonal Frequency Multiple Access (OFDMA) with preamble puncturing [9] to use non-contiguous channels. With all this, given the number of Basic Service Sets (BSSs) and devices in crowded deployments (see Fig. 2), we can say that CB is a problem with a combinatorial action space.

To harness the available spectrum within the CB operation, Dynamic CB (DCB) mechanisms [7, 8] are applied to decide the set of channels for transmitting on a per-packet basis, thus potentially improving performance. In [8], the following DCB policies were proposed and analyzed:

• •

  Static Channel Bonding (SCB): a transmitter is allowed to use the entire set of channels only, thereby limiting the election of any subset of channels. While such a policy may optimize the performance in isolated deployments, it lacks the necessary flexibility to deal with inter-BSS interference.

• •

  Always-Max (AM): in this case, the widest combination of channels is picked upon having sensed them free during the backoff procedure. While such a policy seems to properly harness the available spectrum, it has also been shown to generate starvation and other issues as a result of inter-BSS interactions.

• •

  Probabilistic Uniform (PU): as an alternative to SCB and AM, PU is introduced to add some randomness in the process of picking free channels, so that any combination is chosen with the same probability. This policy has been shown to improve both SCB and AM in some scenarios in which flow starvation was present due to inter-BSS interactions.

To better illustrate the behavior of CB policies, Fig. 3 shows a simplification of the transmission procedure that an STA follows when implementing AM. In particular, CB may be applied over channels 1-4. Based on the AM policy, at time $t_{1}$, the STA can transmit only over channel 2, which is the only one that is sensed free at the moment of initiating a transmission. Similarly, in $t_{2}$, both channels 1 and 2 are found free, so the transmission is performed over those two channels. Finally, provided that the entire spectrum is free, a transmission over channels 1-4 is performed at $t_{3}$. Notice that, if applying SCB, the STA would have not been able to transmit until $t_{3}$. Alternatively, regarding PU, any combination of free channels could have been selected in $t_{2}$ and $t_{3}$.

Through the analysis conducted in [8], it was shown that the right channel choice is not always trivial (i.e., selecting the widest channel does not necessarily entail achieving the highest performance). First of all, using wider channels entails spreading the same transmit power over the selected channel width, which can potentially affect the data rate used for the transmission, and therefore the capabilities of the receiver on decoding data successfully. Moreover, the potential gains of DCB in crowded deployments are hindered by the interactions among Wi-Fi devices, which may provoke contention or collisions. The fact is that WLAN deployments are unplanned and operate under Carrier Sense Multiple Access (CSMA). From the perspective of a given transmitter-receiver pair, such lack of coordination leads to uncontrolled interference that can potentially degrade their performance.

Other DCB mechanisms were proposed in [10, 11, 12, 13], which include collision-detection, carrier sensing adaptation, or traffic load awareness. More recently, ML and game theory have been applied to address CB as an online decision-making problem involving multiple agents [14, 15].

In view of the complexity of selecting the best configuration of channels, the proposed problem statement had the goal of shedding light on the potential role of ML in DCB. In particular, it served to gather participants’ proposals of ML models able to predict the performance of different CB configurations. This information can be used by a decision-making agent to choose the best configuration of channels before initiating a transmission. Throughput prediction in WLANs has been widely adopted for performance analysis through mathematical models, including the well-known Bianchi model [16], Continuous-Time Markov Networks (CTMNs) [17], or stochastic geometry [18]. However, these well-known models lack applicability for online decision-making because they fail to capture important phenomena either on the PHY or the MAC, or they entail a high computational cost. Thus, modeling high-dense complex deployments through these models may be highly inaccurate or simply intractable.

In this regard, we envision ML models to assist the CB decision-making procedure in real-time. The fact is that ML can exploit complex characteristics from data, thus allowing to solve problems that are hard to solve by hand-programming (see, for instance, its success in image recognition). Besides, ML models can be trained offline and then improve their accuracy with measurements acquired online. To the best of our knowledge, this is an under-researched subject. While ML has been applied for predicting aspects related to Wi-Fi networks, such as traffic and location prediction [19, 20], it has been barely applied for explicitly predicting their performance. In this context, the work in [21] provided an ML-based framework for Wi-Fi operation, which includes the application of Deep Learning (DL) for waveforms classification, so that WLAN devices can identify the medium as idle, busy, or jamming. Closer in spirit to our work, [22] proposed an ML-based framework for WLANs’ performance prediction.

To motivate the usage of ML for predicting WLANs’ performance, we provide an open dataset²²2The dataset has been made publicly available at https://zenodo.org/record/4059189, for the sake of openness. obtained with the Komondor simulator.³³3https://github.com/wn-upf/Komondor, Commit: d330ed9. Komondor is an open-source IEEE 802.11ax-oriented simulator, which fundamental operation has been validated against ns-3 in [23]. Komondor was conceived to cost-effectively simulate complex next-generation deployments implementing features such as channel bonding or spatial reuse [24]. Besides, it includes ML agents, which allows simulating the behavior of online learning mechanisms to optimize the operation of WLANs during the simulation.

The dataset generated with Komondor has been used for training and validate ML models in the context of the ITU AI for 5G Challenge. The assets provided comprise both training and test data sets corresponding to multiple random WLAN deployments at which different CB configurations are applied. As for training, two separate enterprise-like scenarios, namely, training1 and training2, have been characterized. In each case, a different fixed number of BSSs coexist in the same area, according to users’ density.

In training1, there are 12 APs, each one with 10 to 20 associated STAs. Regarding training2, it contains 8 APs with 5 to 10 STAs associated with each one. For both training scenarios, three different map sizes have been considered (a, b, and c), where STAs are placed randomly. Similarly to training scenarios, the test dataset includes a set of random deployments depicting multiple CB configurations and network densities. In this case, four different scenarios have been considered according to the number of APs (4, 6, 8, and 10 APs). Note, as well, that, for each type of scenario, 100 and 50 random deployments have been generated for training and test, respectively. In all the cases, downlink UDP traffic was generated in a full-buffer manner (i.e., each transmitter has always packets to be delivered). Table 2 summarizes the entire dataset in terms of the simulated deployments.

With respect to input features, these are included in the files used for simulating each random deployment. In particular, the most relevant information to be used for training ML models is:

1. 1.

   Type of node: indicates whether the node is an AP or an STA.

2. 2.

   BSS id: identifier of the BSS to which the node belongs.

3. 3.

   Node location: {x,y,z} coordinates indicating the position of the node in the map.

4. 4.

   Primary channel: channel at which carrier sensing is performed.

5. 5.

   Channels range: minimum and maximum channels allowed for bonding.

6. 6.

   Transmit power: power used for transmitting frames.

7. 7.

   Sensitivity threshold: used for detecting other transmissions and assess the channels’ availability.

8. 8.

   Received Signal Strength Indicator (RSSI): power detected at receivers from their corresponding transmitters.

9. 9.

   Inter-BSS interference: power sensed from other ongoing transmissions.

10. 10.

    Signal-to-Interference-plus-Noise Ratio (SINR): average SINR experienced during packet receptions.

Regarding output labels, we provide the throughput obtained by each device during the simulation, being the APs’ throughput the aggregate throughput of each BSS (i.e., the sum of all the individual STAs’ throughput in a given BSS). Besides, the airtime per AP is provided, which indicates the percentage of time each BSS has occupied each of its assigned channels.

Wireless networks can be represented by graphs G=(V, E), where V is the set of nodes, i.e., STAs and APs, and E represent wireless links. Typically, DL approaches deal with graph-structured data by processing the data into simpler structures, e.g., vectors. However, nodes and links in high dense WLAN deployments are characterized by a set of high-dimensional features, thus complicating the graph-like structure of the problem and therefore hindering the application of deep learning.

To overcome the problem of data representation, Graph Neural Networks (GNNs) have been proposed as neural networks that operate on graphs intending to achieve relational reasoning and combinatorial generalization [26]. Accordingly, the wireless interactions between STAs and APs (connectivity, interference, among others) can be easily captured via a graph representation.

Therefore, we select a GNN approach to predict the throughput of the devices in a WLAN. In particular, each deployment is considered as a directed graph where STAs and APs are the graph’s nodes. Besides, we define two types of nodes present in the dataset, each one having generic or specific features. For instance, parameters like channel configuration are related to both types of nodes, while SINR is only related to STAs, and airtime is exclusive for APs. Furthermore, the edges are defined based on the type of wireless interaction derived from the dataset. We consider two types of interactions, namely AP-AP interactions (represented by the interference map), and AP-STA interactions (represented by the RSSI values). For completeness, we define an additional edge feature based on the distance of every transmitter-receiver pair. The features considered for training the proposed GNN are summarized in Table 3.

Concerning the GNN model, we have used an implementation of the Graph Network Block (GNB), as proposed in [27]. A GNB contains three update functions and three aggregation functions where the computation is done from edge to nodes, and then to global parameters. So first, the edge’s features are updated and aggregated into the node features, then the node features are updated having into account the vicinity within depth/range defined by the number of GNBs, and lastly, the global parameters are updated according to the state of the nodes. Our model follows a layered approach, similar to DL, where each layer is a GNB. The input of the model is a graph representing the deployment, and the output is the predicted throughput of the devices in that deployment. A general overview of our model’s architecture is shown in Fig. 4.

The implementation of the GNB is referenced as a MetaLayer in PyTorch Geometric [28], a geometric deep learning extension library for PyTorch. We defined an edge model that uses two dense layers using Rectified Linear Unit (ReLU) as an activation function in a typical Multi-Layer Perceptron (MLP) configuration. A node model is also defined by two MLPs, one for aggregating the edge features into the node features and the second to update nodes’ states based on their neighbors. Note that, following the formal definition of a GNB in [27], global parameters were not used. Our implementation is available in GitHub [29].

To train the model, we have considered splitting the provided dataset (80% for training and 20% for validation). As each deployment is considered to be represented by a graph, 480 graphs have been used for training, and 120 for validating purposes. The loss function considered to assess the performance of our model is the Root Mean Squared Error (RMSE). The error obtained across all the predictions is used to compare the accuracy of our model’s predictions to the actual results. However, initial results showed that the model mostly focused on predicting the throughput of the APs, given that the error is minimized on large values. Therefore, we proposed a masked loss provided that AP’s throughput should be equal to the sum of the associated STAs’ throughput. The RMSE is calculated using the STAs’ predicted and the APs’ computed throughput.

To address the throughput prediction problem in CB-compliant WLANs, we propose a deep neural network where the information of each BSS is processed independently, thus following the idea of Multi-Layer Perceptron (MLP) [30]. More specifically, the proposed model is a feed-forward deep learning algorithm implemented in python with the support of the PyTorch libraries.⁴⁴4The code used to implement the method proposed by Ramon Vallés is open-access [31].

Our model aims to predict the aggregate throughput of each BSS, rather than the individual throughput at STAs. The fact is that predicting the throughput per STA is very challenging because of the dynamic channel bonding policy used in complex scenarios, which contributes to generating multiple interactions among nodes that cannot be captured at a glance. Accordingly, to derive an overall representation of each BSS, the features from individual STAs are preprocessed so that we consider only their global distribution (mean, and standard deviation). To do so, our model is divided into three main blocks performing different tasks:

1. 1.

   Signal quality: The first block is meant to abstract the interactions among APs and STAs in the RF domain. To achieve this, we consider two separate layers, which process the RSSI, SINR, distance among nodes, and SINR. A parametric Rectified Linear Unit (PReLU) activation function is used together with 1-dimensional batch normalization.

2. 2.

   AP bandwidth: The second block analyzes the available bandwidth of the APs, and it consists of a single linear layer, which receives as input a vector with the corresponding airtime for all the available channels. This layer outputs a 3-dimensional array and is activated with a PReLU function.

3. 3.

   Output: Finally, the last block takes the output from both Signal quality and AP bandwidth blocks and computes the final prediction value. For this final layer, we employ a simple ReLU (instead of a PReLU) to avoid negative throughput predictions.

With this structure, we have built a much more efficient model than if we had used a fully-connected NN with all the STA features. Notice that the proposed model needs far fewer neurons (and thus, less computational force) to capture the most relevant information of each scenario. In our opinion, an excess of neurons in such a complex scenario would result in overfitting, thus making the model less accurate for predicting the performance of new deployments.

As for training the MLP model, we have considered the following key features: the type of device (AP or STA), its location, the minimum, and maximum channel through which it is allowed to transmit, the RSSI, the SINR, and the airtime. The training was performed using 80% of the dataset (keeping the 20% left for validation), following an evaluation criterion based on the RMSE of the predicted throughput with respect to the real value. The Adam optimizer has been used to optimize the training process, which is straightforward to implement, computationally efficient, and memory-efficient. For the training phase, several experiments had been made by modulating the hyper-parameters. The best results were achieved with a learning rate of 0,025 and a total of 700 epochs.

Our proposal includes popular ML regression algorithms such as MLP, Support Vector Machine (SVM), Random Forest, and eXtreme Gradient Boosting (XGboost). These algorithms are backed by rich research, known to do well on regression problems, ease of implementation and deployment, which are important characteristics from the business perspective.

From all the abovementioned methods, XGBoost [32] was selected for competing in the challenge because it offered the highest performance on both the training and validation stages compared with the other models. XGboost is a gradient boosting framework available for multiple platforms, thus providing high portability.

To train our model, we have analyzed the variance of the variables in the dataset, and discarded the features with low or moderate variability. Some of the considered features are the position of nodes, the primary channel, the distance among nodes, or the power received by neighboring devices. To pre-process the selected features, we have applied Yeo-Johnson transformation [33] and normalization. These steps were done for predictors to improve their utility in the models.

The training dataset was split into training and validation, so that we could train the model on the whole training dataset using a 10-fold cross-validation procedure. Further, we used 100 and 300 combinations, respectively, with 10-fold cross-validation.

As for the hyperparameter setting (e.g., max depth or minimum child weight), we tuned hyperparameters using a grid search. More specifically, the hyperparameter values were set using Latin Hypercube Sampling, while their maximum and lower range value for each hyperparameter were mostly pre-determined using default values from tidymodels [34].

Finally, significant efforts have been put to deploy our model. In particular, we have used docker to make our model easy to (re)train and deploy. All the code and documentation has been made publicly available [35].

We formulate throughput forecasting in WLANs as a linear regressing problem, which assumes a linear relationship between the features and label(s) of a given data set {$y_{i}$, $x_{i,1}$, $x_{i,2}$, …, $x_{i,N}$}, and a set of unknown parameters $w$ to be learned (being $w_{0}$ the bias).

Our solution (named Gossip) is based on a linear regression method, and it aims to predict the throughput of an STA in a given WLAN deployment where CB is applied [36]. Based on STAs’ individual throughput, we derive the performance of each AP by aggregating the values of their associated STAs. In particular, Gossip derives the unknown bias and weight parameters $w$ by (i) processing the WLAN dataset, and (ii) applying a neural network to perform regression.

As for the processing part, Gossip takes the input features generated by the Komondor simulator, and selects/generates the most relevant ones: the position of the STA, the AP to which the STA is associated, the RSSI, the SINR, the set of nodes using the same primary channel, and the set of allowed channels. After processing the features, each STA of every deployment is characterized by a feature vector ($x_{i}$,…,$x_{21+3k}$ ), with $k$ denoting the number of wireless channels. Note, as well, that the entire dataset is considered for training, thus combing STAs from different deployments. The rationale is that features such as the number of neighbors in the primary channel, the SINR, and the interference should differentiate STAs from different deployments.

When it comes to the regression problem, Gossip uses a feed-forward neural network with four layers. The input layers pass the input features to two fully connected layers of neurons with a ReLU activation unit. Finally, a single neuron receives the output of the hidden layers and generates the prediction of the throughput. As a remark, the last neuron has a linear activation. It is important to remark that the proposed neural network is mostly meant to tackle linear regression problems. Nevertheless, even if the throughput prediction problem for WLANs is not linear, we expect our model to properly identify local minimum/maximum points that allow providing reasonably prediction results.

To train the proposed neural network, we have used the RMSprop gradient descend method, considering the Mean Squared Error (MSE) as a loss function. Moreover, 50 training episodes and a batch size of 50 STAs have been considered. Thanks to Gossip design, the training dataset is populated with every STA of every deployment present among all scenarios.

To address the objective of predicting the throughput of APs and STAs in typical dense environments, we explore a set of popular regression techniques. With the help of these techniques, we aim to build complex mathematical relationships among features and labels from the dataset, so that performance of WLANs can be predicted at unseen deployments. In particular, we propose using the following techniques:⁵⁵5The code used to implement all the proposed methods by Net Intels is available in Github [37].

1. 1.

   Artificial Neural Network (ANN): ANN method is selected chiefly due to its potential and versatility to model nonlinear and complex relationships in OBSS data elegantly. The proposed ANN is built using Tensorflow and Keras libraries in python [38]. The NN model is designed with one input layer, 7 hidden layers, and 1 output layer (see Fig. 5). The ReLU function is employed to activate hidden layers. In each of the first six hidden layers, there are 1024 nodes. For the seventh hidden layer, there are 512 nodes. The model is trained using Adam optimizer. The batch size and number of epochs for training, after multiple trials, were set to 250 and 1000 respectively.

2. 2.

   K-Nearest Neighbor (KNN) regression: For OBSSs involving several AP-STA combinations, the dynamics of interrelations between entities of OBSS rely predominantly on the relative positioning of AP/STAs. To abstract such complexity in a cost-effective manner, KNN is selected, which is characterized by its simplicity, speed, and protection against high variance and bias. The KNN model is built using the Scikit Learn library in python [39]. The inbuilt KNearestRegressor function is directly used, which neighbor number is fixed to 10. The algorithm for structuring the k-dimensional space of the dataset (Ball Tree, KDTree, or Brute Force) is automatically selected based on input values.

3. 3.

   Random Forest Regression: Motivated by the fact that the interrelationship between the features is non-linear, we propose dividing the dataset dimensional space into smaller subspaces. To generalize the data and for better feature importance, an ensemble of trees forming a Random forest is used. Random forest mechanisms are useful to reduce the spread and diversion of predictions. The proposed Random Forest regression is built using Scikit Learn, an ensemble module of the Sklearn library. The default number of trees was set to 100, which split is performed according to the mean squared error function. The maximum depth of the tree is set to 10.

For all the proposed methods, we have first preprocessed the dataset comprising six hundred different random deployments. In particular, static features such as the Contention Window (CW) were not included for training purposes. As for the rest of the features, the P-values of each feature were considered to identify the ones with higher significance. To firm up the choice of features to be selected for training, we have used the ones with a higher correlation degree (see Fig. 6).

Based on this analysis, we decided to include the following features: X and Y coordinates, primary channel, minimum and maximum channel allowed, SINR, and RSSI. The features data is standard scaled before being fed into the regression models. Only data for STAs (stations) are considered for training and the throughput values of STAs are predicted using the models. The sum of the throughputs of STAs gives the throughput of the respective AP. For training purposes of all the three methods, the data is split (80% for training and 20% for validation).

With contributions from participants around the globe, the ITU AI for 5G Challenge has unprecedentedly established a platform for addressing important problems in communications through ML. As for the performance prediction problem in CB WLANs (referred to as PS-013), the challenge has allowed us to glimpse the potential of ML models for addressing it.

This paper provides a compendium of ML models proposed for throughput prediction in WLANs, including popular models such as neural networks, linear regression, or random forests. In particular, we have overviewed the proposed models and analyzed their performance in the context of the challenge. By opening the dataset used during the competition, we encourage the development of mechanisms that improve the baseline performance shown by the ML models presented in this work.

From the performance evaluation done in this paper, we have drawn the following conclusions:

1. 1.

   First, even if the dataset was not particularly big,⁶⁶6Youtube-8M Dataset (http://research.google.com/youtube8m/) consists of 350.000 hours of video, while only six hundred different random deployments have been used in this paper for training purposes. some of the proposed ML models achieved good results. This is quite a positive result since it opens the door to ML models that can be (re)trained fast, thus becoming suitable for (near)real-time solutions.

2. 2.

   Second, most of the proposed DL-based models have shown higher accuracy for the denser and more complex deployments (which more closely match the training scenarios) than for the sparser ones. While capturing complex situations is quite a positive result, the pitfalls observed in simpler deployments also suggest that out-of-the-box DL methods may fail at capturing the relationship between interference and performance of WLANs. In this regard, well-known models characterizing WLANs (e.g., SINR-based models [40]) can potentially be incorporated into the ML operation for the sake of improving accuracy, thus leading to hybrid model-based and data-driven mechanisms.

3. 3.

   Third, and related to the previous point, GNNs have been shown to be particularly useful to capture the complex interactions among devices in WLANs, both in terms of interference and neighboring activity. In particular, we have realized the importance of pre-processing the dataset in order to obtain accurate prediction results. Deriving information specific to the problem (i.e., signal quality, interference) has turned out to be essential for the sake of generalization.

4. 4.

   Finally, we remark the importance of cost-effectively predicting the performance in WLANs, which may open the door to novel mechanisms using these predictions as heuristics for online optimization. The incorporation of these kinds of models to WLANs is expected to be enabled by ML-aware architectural solutions [41].

The availability of data for training is key for the success of ML application to future 5G/6G networks. Given the current limitations in acquiring data from real networks, simulators emerge as a practical solution to generate complementary synthetic data for training ML models. The fact is that data may be scarce because, among other reasons, measurement campaigns are costly, data from networks involves privacy concerns, or data tenants are not willing to share their data.

Through the problem statement discussed in this paper, we have contributed to showcase how synthetic data can be used to train ML models for networks. A notorious advantage is that network simulators allow characterizing complex deployments, sometimes representing unknown situations, so they help to train and validate ML models.

Beyond generating data for training, network simulators are envisioned to serve as secure platforms for testing, training, and evaluating ML models before being applied to operative networks [42]. In consequence, we foresee the adoption of simulators into future ML-aware networks as a key milestone for enhancing both reliability and trustworthiness in ML mechanisms.