Intelligent hybrid automatic repeat request retransmission for multi-band Wi-Fi networks

With the increase of Wi-Fi connections and throughput, the problems posed by unreliable connection and unstable delay need to be urgently solved through advanced methods. The IEEE 802.11ax standard was designed to improve the reliability of connection and increase the throughput in dense Wi-Fi scenario where the inter-site interference becomes more serious and the spectral resources are more insufﬁcient. Furthermore, in the next generation of Wi-Fi standard IEEE 802.11be, the hybrid automatic repeat request (HARQ) and multi-band technologies will be introduced to solve the problems of unstable delay, serious interference and insufﬁcient spectrum resources. In this paper, retransmission schemes combined multi-band and hybrid automatic repeat request are deeply studied to improve the retransmission efﬁciency. Combining hybrid automatic repeat request with multi-band, this paper proposes three kinds of retransmission mode: insisting on the current frequency band (ICFB) retransmission, switching to the backup frequency band (SBFB) retransmission and concurrent retransmitting on dual-band (CRDB). Then, an intelligent selection algorithm of retransmission mode based on machine learning is designed to determine the optimal mode. Theoretical analysis and experimental results show that the proposed method can greatly improve the


INTRODUCTION
Recently, Wi-Fi has become an indispensable mobile service access solution. Applications such as smart homes and augmented virtual reality raise urgent demands on Wi-Fi devices for lower-latency and more-reliable connection. However, since Wi-Fi operates on the unlicensed 2.4/5 GHz frequency bands, there exists significant signal interference and collision during competitive accesses, particularly in dense Wi-Fi networks [1], which will increase the probability of data transmission failure. Therefore a good retransmission mechanism will play an important role in improving the transmission efficiency of Wi-Fi networks. Nevertheless, IEEE802.11ax based Wi-Fi still uses the automatic repeat request (ARQ) technique that works in single frequency band to correct transmission errors [2]. In the process of ARQ, if the received data cannot be correctly decoded, the data will be discarded and requested to be retransmitted until the number of retransmission attempts reaches the maximum limit of seven or the data are correctly decoded [3], which obviously results in low retransmission efficiency. As a mature and efficient retransmission technique, the hybrid automatic repeat request (HARQ) had been proved effective in cellular networks, especially for edge users. Recently, HARQ is also considered as one of the key technologies of the next generation Wi-Fi standard IEEE 802.11be in order to meet the requirements of reliable connection and large throughput [4]. Technically, HARQ combines forward error correction (FEC) with ARQ and adopts soft combining mechanism, which merges the retransmitted data with the previously incorrectlydecoded data [5,6]. It effectively utilizes the useful information in the previously received data achieving a diversity gain and thereby increasing the retransmission successful probability.
Extensive tests and studies have shown that the service loads on 2.4 and 5 GHz bands are not balanced especially in the dense Wi-Fi networks. For example, the 5 GHz band with more spectral resources may bear more service load than 2.4 GHz [7], which may cause high collisions in the 5 GHz band. Based on this observation, an HARQ mechanism based on multi-band is proposed in this paper, which can reasonably dispatch the retransmission data to relatively less busy band to balance the load and reduce the retransmission times.
The main contributions of this paper are as follows: 1. To enhance the retransmission efficiency, an HARQ mechanism is designed based on multi-band, including the design of communication mode and signalling. 2. To adapt to different retransmission situations, three retransmission modes are proposed, i.e. insisting on the current frequency band (ICFB), switching to the backup frequency band (SBFB) and concurrent retransmitting on dual-band (CRDB). 3. To implement intelligent decision of retransmission, a selection algorithm of retransmission mode based on machine learning is proposed.
The rest of this paper is organized as follows. In Section II, the technical characteristics of 802.11ax protocol are investigated and an HARQ mechanism is designed based on multiband, including the design of three kinds of communication mode and signalling. In Section III, the selection algorithm of retransmission mode based on machine learning is presented and discussed in details. In Section IV, the mathematical and simulated models are formulated to illustrate the proposed schemes and simulation results are presented to validate the feasibility of the proposed schemes. Finally, Section V concludes this paper.

SYSTEM MODEL
In this section, we introduce our proposed multi-band HARQ retransmission scheme. Without loss of generality, we take the HARQ chase combine (HARQ-CC) as a study baseline [8,9], where the contents in each retransmission are the same. The main idea of our scheme is to dynamically allocate the retransmissions on different frequency bands. Accordingly, we design three strategies, namely ICFB, SBFB and CRDB, whose details are as follows.

Insisting on the current frequency band
Operating at 5 GHz band, IEEE 802.11ax has two multiple access modes, i.e. carrier sense multiple access with collision detection (CSMA/CD) based free competition and centrally-controlled scheduling. Uplink-OFDMA random access (UORA) in 802.11ax employs orthogonal frequency division multiple access (OFDMA) technology, which divides the frequency resources into multiple resource units (RUs), and then the station (STA) can transmit in multi-link through competing RU resources in different frequency bands. In downlink (DL), after collecting the transmission requests from STAs, the access point (AP) allocates RUs for each STA [10]. The Scheduling-based mechanism specific scheduling process is shown in Figure 1. Obviously, this mechanism requires AP to frequently collect the transmission requests from STA, which will result in signalling overhead, especially when considering the retransmissions. In order to solve this problem, an RU reservation mechanism is adopted in IEEE 802.11ax [11]. As shown in Figure 2, being assigned with an RU, STA n 1 appends the RU reservation request in PPDU MAC header to re-use the same RU resource in the next TF-R phase. After successfully receiving PPDUs from an STA, the AP replies multiple-block-ACK (M-BA) for confirmation. In addition, the number of reserved RUs, i.e. the reserved order, is denoted as m. When a new TF-R frame contains the information of the reserved RU, n 1 can directly transmit its UL PPDU on the specified RUs without contention. Since in IEEE 802.11ax STAs can actively request RU resources instead of passively waiting for scheduling information from AP, we propose the ICFB mode as shown in Figure 3, where RUs can be reserved actively by the initiator of transmission when the received data are failed to decode. After successful decoding, the receiver resets the reserved order of RUs and releases the rest of RU resources.

Switching to the backup frequency band
With the popularity of dual-frequency Wi-Fi technology, more and more equipments support both 2.4 and 5 GHz frequency bands in Wi-Fi networks [12]. Although dual-band increases the bandwidth of transmission, there is an imbalance in the distribution of service load between the two bands in actual Wi-Fi networks [13]. For example, the increase of transmission pressure in 5 GHz frequency band with wider bandwidth will result in a shortage of its resources and an increased probability of congestion and collision, which then increases the retransmission rate. However, if there are relatively few STAs accessing 2.4 GHz frequency band or less interference in this band, backing up the retransmitted data in 5 GHz frequency band to 2.4 GHz frequency band can reduce the traffic load of 5 GHz frequency GHz frequency band is serious, the retransmitted data can also be backed up to the 5 GHz frequency band to reduce the traffic load of 2.4 GHz frequency band and improve the retransmission efficiency, as shown in Figure 4. As we know, on-channel tunnelling (OCT) operation defined by IEEE 802.11ax enables STA to transmit multiple MPDUs (MMPDUs) constructed by different frequency band of the same device. As shown in Figure 5, a system management entity (SME) of a multi-band device may instruct its MAC layer management entity (MLME) to communicate with MLME of a peer multi-band device using OCT services [14]. Management frames and on-channel tunnel request/report frames transmitted in OCT frames between peer MLME entities are used for STA/AP to encapsulate MMPDUs in order to transmit the MMPDUs to MLME of peer STA/AP within the same multiband device. If SBFB mode is needed, OCT frames designed for retransmission are used to exchange the frequency band information and apply for channel resources to the backup frequency band. The designed frame structure is shown in Figure 6.
After applying for resources from the backup frequency band and packaging the retransmission data in OCT frames, the retransmission initiator will initiate a fast session transfer (FST) application, as shown in Figure 7. In IEEE 802.11ax, devices with multi-band may work on different frequency bands concurrently or non-concurrently. If the two frequency bands work non-concurrently, the session is going to be quickly transferred from the current frequency band to the backup frequency band, and it will backoff to the source frequency band once the retransmission is completed. If the two frequency bands work concurrently, the backup frequency band is used for retransmission and the backup frequency band resources are released once the retransmission is completed.

Concurrent retransmission on dual-band
Different from RBFB, the characteristic of CRDB mode is that the source frequency band still works and continues to retransmit the incorrectly-decoded data, after applying for retransmission data in backup frequency band until receiving ACK from the receiving end, as shown in Figure 8. When the retransmission is completed, the backup frequency band will be released directly and the current frequency band still works. It is worth mentioning that although the CRDB can greatly improve the retransmission efficiency, it occupies the resources of two frequency bands and generates large overhead.

INTELLIGENT SELECTION ALGORITHM FOR RETRANSMISSION
In actual Wi-Fi networks, there are various reasons for transmission failure, such as sudden deterioration of channel quality or collision interference. However, the cause of current transmission failure cannot be easily identified and it is difficult to select the optimal retransmission mode based on the cause of transmission failure. Therefore, we propose a selection algorithm of retransmission mode based on machine learning method, which intelligently selects the most appropriate retransmission mode through analysing the characteristics of network scenes.

Probability of successful transmission
In HARQ mechanism, every retransmitted datum will be combined with the previous erroneous datum. As a result, the where Γ(a) represents the gamma function with parameter a. The curve of incomplete gamma function is characterized as that the value of Γ(x, a) approaches 1 as x approaching infinity for any a, which completely accords with the characteristic of HARQ when performing multiple retransmissions. Therefore, combining the incomplete gamma function with the characteristics of HARQ can obtain the function of probability of successful transmissions.
where Γ inc is an incomplete gamma function, z is the retransmission times, m is the Nyquist fading factor, B(r ) is the bit error rate of current transmission, r 0 = 2 R − 1, R is the symbol rate andr is the average signal-to-noise ratio of current channel. When the retransmission data need to switch to a backup frequency band, the probability of successful transmissions model can be written as: where z 1 and z 2 represent the retransmission times in two frequency bands respectively.

Algorithm model
Our selection algorithm of retransmission mode based on machine learning designed is divided into two parts, namely the offline stage and the online stage, which is shown in Figure 9. The offline stage is a process of unsupervised clustering, while the online stage is a process of supervised deep learning. The clustering searches for the distribution of the channel through the channel state information of the Wi-Fi networks, which is the precursor of the deep learning. The process of deep learning takes the results of clustering to train model that makes the decision of the retransmission mode. This algorithm defines the basic set of data before clustering in Wi-Fi as x i = ( 1 , 2 , … , 7 ), where 1 − 7 represent the retransmission times, the SINR of channels, the channel occupancy ratio in 2.4 GHz, the channel occupancy ratio in 5 GHz, the state of switching, derivative of throughput in 2.4/5 GHz and the average delay of the current link respectively. Furthermore, the theoretical probability of successful transmissions is taken as an additional auxiliary parameter of the model to enhance the rationality in selecting retransmission mode.
The traditional K-means clustering algorithm randomly selects k samples as the initial mean vector ( 1 , 2 , … , 7 )and takes ( 1 , 2 , … , 7 )as the clustering centroids, then calculates the distance from each sample in the sample set D = {x 1 , x 2 , … , x m } to the k clustering centroids respectively and divides each sample into the nearest clustering centroids. Eventually the algorithm recalculates the centroid of each cluster and repeatedly searches for the new centroid cluster until the centroid position no longer changes or reaches the maximum number of iteration rounds [15,16]. However, due to the problem of local optimal of the K-means algorithm and the low convexity of the channel state information of Wi-Fi networks, it is difficult to make the optimal selection of retransmission mode. In order to increase the convexity of channel state information and improve the clustering efficiency, we herein design an optimized K-means algorithm. The improved algorithm weights each vector when calculating the distance from each sample point in sample set D = {x 1 , x 2 , … , x m } to k clustering centres respectively.
where = ( 1 , 2 , … , 7 ) is the weighted vector of the input sample, which is used for balancing the characteristics of the input data. Thus, the optimized algorithm can be described as Table 1.
The classified feature data after clustering are taken as the training set and the verification set of the neural network. Since the input data of neural network have been extracted by clustering method, this paper takes a relatively mature inverse error back propagation (BP) algorithm [17,18]. The designed inverse error propagation neural network is shown in Figure 10. The input layer includes 7 neurons of channel state information, and one neurons of probability of successful transmissions. The hidden layer is full connection layer to increase the network depth, and the output is softmax layer. The sigmoid function of softmax layer is written as:  Therefore, the decision of retransmission mode is to select the one with the largest value from the output.

Simulation model
The simulation modelling and analysis of selection algorithm based on machine learning are carried out on MATLAB. In the simulation process, it is assumed that the state of each channel in all frequency bands of the Wi-Fi network is known, and the interference from other networks is ignored. The simulation parameters are shown in Table 2.
The simulation observes the improvement of the algorithm respect to delay, average throughput and average retransmission times. For traditional HARQ, when the 2.4 or 5 GHz is operated in non-competition mode, the latency after z times retransmission can be denoted as: where W i represents the window value of the first transmission contention and BA represents the delay of block ACK. In addition, if the backup frequency band is required and the 5 GHz frequency band is operating in the scheduling state, the delay after z times retransmission can be denoted as Equation (7), where z 1 and z 2 represent the retransmission times in the current frequency band and the backup frequency band respectively. RU ave. , Tri. and MBA represent the average length of RU, the length of trigger frame and the length of MU-Block ACK frame respectively, and FST refers to the delay caused by FST. In addition, when the 5 GHz frequency band also works in the state of free competition, the transmission delay can be Input: Sample set D = {x 1 , x 2 , … , x m } Number of clustering cluster k(k = 3) Mean weight matrix = ( 1 , 2 , … , 7 ) Process: 1: Randomly select k samples as the initial sample mean vector ( 1 , 2 , … , 7 ) 2: Repeat 3: Order C i = ∅(1 ≤ i ≤ k) 4: For j = 1, 2, 3, … , m + 1 do 5: Calculate the weighted distance between x i and (1 ≤ i ≤ k): Determine the cluster mark of x j : ie. (1,2,…,k) ] 7: Divide sample x j into the corresponding cluster: C i = C i ∪ {x j } 8: End for 9: For i = 1, 2, … , k do 10: Calculate the new mean vector: If ′ i ≠ i then 12: Update the current mean vector i to ′ In our algorithm, the throughput of Wi-Fi network is an important criterion to measure its performance. For the sake of simplicity of modelling, the throughput is defined as the average throughput that the ratio of the transmitted data length to the Time synchronization and channel estimation Idealization time interval at which the sending end receives the ACK. Then the throughput in the backup frequency band and the throughput in the current frequency band are calculated as S 1 and S 2 (see Equation (10)) respectively.
The selection algorithm of retransmission mode mainly involves two parts of algorithm implementation, unsupervised clustering states of the Wi-Fi network and supervised deep learning decisions. The K-means algorithm divides the characteristics of different retransmission modes into clusters C = {c 1 , c 2 , … , c 7 } for a given sample set D = {x 1 , x 2 , … , x m }, and obtains the minimum variance of squares: x is the mean vector of the cluster C i . The smaller the mean-square-error of the above Equation (9), the higher the similarity of samples within the cluster. In addition, the most important step of the BP algorithm is to determine the parameters of the neural network. In order to get the best classify retransmission mode, this simulation sets different learning rates ∈ (0, 1) to train the neural network. According to the clustering result, the sample set duplication D = {x 1 , x 2 , … , x m } can be written asD = {y 1 , y 2 , … , y m }, where y i represents the best retransmission mode to which the sample belongs. As can be seen from the neural network structure described above, assuming that the output of training sequence (x i , y i ) isŷ i = {ŷ 1 ,ŷ 2 ,ŷ i3 }, the mean-square-error of the network on the sam-

FIGURE 11
Clustering results without weighted processing ple can be written as: when the mean-square-error is obtained, the parameters of the neural network are adjusted in the direction of the negative gradient of the target based on the strategy of gradient descent, and the updated values of the weighted matrix of ji , a ji and b ji are obtained. Finally, the network performance is evaluated by means of tenfold cross validation.

Performance analysis
In order to verify the performance of the algorithm and the accuracy of the theoretical analysis, a network with a maximum number of 50 STAs is considered in this paper. The STA calibration is set to different requirements of transmission, including voice and game that need a stable delay, and video browsing that need large throughput. Further, with the increasing number of STAs, the interference strength values of different frequency bands and different RU in Wi-Fi network is adjusted randomly and dynamically. The simulation results show that the optimized K-means algorithm is powerful to select the best retransmission mode.
The basic data set of Wi-Fi network is iterated in K-means algorithm (K = 3) for each round. Due to the irregular input data of Wi-Fi network status, in the simulation process, with the infinite increase of iteration rounds, it is difficult to produce exactly the same iteration results. Therefore, the quotient of the sample mean-vector of the current iteration and the sample mean of the previous iteration is taken as the end of the algorithm. It can be seen that by using the optimized K-means algorithm, when the input data are weighted, the number of iteration rounds is significantly reduced, and the clustering effect is getting better, as shown in Figures 11  and 12. In the optimized K-means algorithm, the weighted vector will affect the clustering performance. For example, if 1 increases the weight of channel occupancy, the mean-squareerror of clustering will decrease more rapidly as the number of iterations increases. However, 2 can obtain a relatively lower mean-square-error by increasing the weight of throughput derivative, but the gradient of mean-square-error will be smaller as the number of iterations increases. On the whole, setting a reasonable weighted vector can effectively improve the clustering efficiency and reduce the iteration times, as shown in Figure 13.
Although the optimized K-means algorithm can well improve the average throughput of the system and reduce the delay of retransmission, the average throughput does not increase by an order of magnitude, as shown in Figure 14. In summary, the average delay in the optimized K-means algorithm has a great performance improvement compared with the simple retransmission mechanism, but the peak time delay of STA is still unstable, as shown in Figure 15. After the process of unsupervised deep learning the best retransmission mode is corrected by BP algorithm, which greatly reduces the average delay and keeps the maximum delay within 10 ms of the system. In addition, through the simulation results of the algorithm, as shown in the Figure 16, the error rate of the algorithm can be controlled within 0.01 when the neural network is trained by tenfold cross validation.

CONCLUSION
This paper combines HARQ with multi-band technology in the next Wi-Fi standard IEEE802.11be and designs an HARQ retransmission scheme based on multi-band. A selection algorithm of retransmission strategy based on machine learning is proposed by taking the advantages of K-means and BP algorithms. The simulation results show that the method of machine learning can effectively improve the average

FIGURE 16
Error probability of the mode selection algorithm throughput and greatly reduce the average and maximum delay of the dense Wi-Fi networks. It is worth mentioning that, the algorithm designed in this paper only considers the 2.4 and 5 GHz frequency bands, without considering the 6 GHz unlicensed frequency band that will be used by the next generation of Wi-Fi [19,20]. If the three frequency bands all are considered, it can be predicted that the retransmission efficiency and the throughput of Wi-Fi networks could be further improved.