Received 21 September 2018; revised 25 December 2018; accepted 11 February 2019.

Date of publication 14 February 2019; date of current version 16 September 2021.

Digital Object Identifier 10.1109/TETC.2019.2899407

An Intelligent Route Computation Approach Based

on Real-Time Deep Learning Strategy for Software
Defined Communication Systems
BOMIN MAO , (Student Member, IEEE), FENGXIAO TANG , (Student Member, IEEE),
ZUBAIR MD. FADLULLAH , (Senior Member, IEEE), AND NEI KATO , (Fellow, IEEE)
The authors are with the Graduate School of Information Sciences, Tohoku University, Sendai 980-8577, Japan
CORRESPONDING AUTHOR: F. TANG ([email protected])

ABSTRACT Software Defined Networking (SDN) is regarded as the next generation paradigm as it simpli-
fies the structure of the data plane and improves the resource utilization. However, in current Software
Defined Communication Systems (SDCSs), the maximum or minimum metric value based routing strategies
come from traditional networks, which lack the ability of self-adaptation and do not efficiently utilize the
computation resource in the controllers. To solve these problems, in this paper, we utilize the deep learning
technique to conduct the routing computation for the SDCSs. Specifically, in our proposal, the considered
Convolutional Neural Networks (CNNs) are adopted to intelligently compute the paths according to the input
real-time traffic traces. To reduce the computation overhead of the central controller and improve the adapta-
tion of CNNs to the changing traffic pattern, we consider an online training manner. Analysis shows that the
computation complexity can be significantly reduced through the online training manner. Moreover, the sim-
ulation results demonstrate that our proposed CNNs are able to compute the appropriate paths combinations
with high accuracy. Furthermore, the adopted periodical retraining enables the deep learning structures to
adapt to the traffic changes.
INDEX TERMS Software defined communication systems, deep learning, real-time learning, routing
computation

I. INTRODUCTION architectures to fulfill management in high efficiency, in the

In recent decades, the computation platforms have experi- SDNs, the complex network control logic is separated to the
enced several times of exponential increase of computation central controllers which consist of various computation plat-
capacities because of the revolutionary improvement in the forms [6], [7]. Similar to the cloud-based computing applica-
hardware architecture and the manufacturing technology. tions, the controllers conduct all the computation tasks for
Inspired by the development of computation platforms as the switches [8]. Moreover, the general architecture of the
well as their decreasing cost, various kinds of services aiming SDN controllers allows the upgrade of network management
at offering companies and users computation assistance are algorithms fulfilled by updating the corresponding applica-
booming, such as the cloud computation. Moreover, the tions, which is more flexible than conventional networks.
Internet Service Providers (ISPs) apply these hardware to As shown in Figure 1, the structure of SDN consists of three
construct various kinds of wireless networks, e.g., Fiber- planes: the data plane, the control plane, and the application
Wireless (FiWi), Device to Device (D2D), and 5G, to satisfy plane. The control logic of the whole network is implemented
users’ requirements in different scenarios [1]—[3]. Since in the control plane while the switches in the data plane are just
these networks have various infrastructures and topologies, responsible for forwarding packets. Any new network man-
to solve the difficulty in managing all these networks, agement application can be installed in the controller thro-
researchers proposed the Software Defined Networking ugh the Northbound Application Programming Interfaces
(SDN) [4], [5]. Different from the traditional networks which (NAPIs) [4], while the communications between the controller
integrate the algorithms into the proprietary hardware and switches are fulfilled via the Southbound Application

2168-6750 © 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE
1554 permission.
Authorized licensed use limited to: Mepco Schlenk See ht_tps://www.ieee.org/publications/rights/index.html
Engineering College. Downloaded on October 09,2024 for more
at information. VOLUME
05:53:06 UTC from IEEE 9, NO.
Xplore. 3, JULY-SEPT.
Restrictions 2021
apply.
Mao et al.: An Intelligent Route Computation Approach Based on Real-Time Deep Learning Strategy for Software Defined Communication Systems

controller to initialize the utilized Convolutional Neural Net-

works (CNNs). Then, in the running phase, the CNNs are
adopted in the controller to choose paths. Furthermore, to
adapt the trained CNNs to the changing traffic patterns as
well as reduce the training computation overhead, in the run-
ning phase, the switches keep recording the traffic trace for
periodically retraining the CNNs in the controller. The major
contributions of this paper can be summarized as follows.

First, we analyze the problem of conventional routing
protocol for the SDCSs and explain the reasons.

Second, we adopt the deep leaning technique to con-
duct the routing path computation. The proposed deep
learning structures can learn from the previous experi-
ence, which enables the controller to become intelligent
FIGURE 1. The SDN architecture.
to avoid the repetitions of wrong decisions.

Last but not the least, the online fashion is adopted to
Programming Interfaces (SAPIs) [9], [10]. It can be found that train our proposed CNNs. By this way, the computation
besides the improved feasibility in SDN, the structures of overhead can be reduced and the considered CNNs can
routers and switches get significantly simplified and unified adapt to new network surroundings automatically with-
due to the separation of complicated network logic. out manual intervention. Moreover, this method over-
To fit the new network scenario as well as the improved comes the limitation of supervised learning that a large
computation capacity in SDN, the network algorithms should volume of labeled data are not always available.
have been updated or redesigned. However, the packet for- The remainder of the paper is organized as follows.
warding algorithms in current Software Defined Communi- Section II first describes the previous work on the deep learn-
cation Systems (SDCSs) [11] still follow the conventional ing based studies in the field of communication networks.
manner, which is based on the maximum or minimum metric Section III presents preliminaries on SDN and CNNs. In
values [12], such as the Shortest Path (SP) algorithm. Since Section IV, we describe the problems in the routing strategies
the paths are computed according to fixed rules, when similar for current SDCSs, and then discuss our proposal to over-
traffic patterns happen, the controller chooses the same paths come the problems. The detailed procedures of our proposal
even the decision has been previously proved wrong, which are presented in Section V. Section VI analyzes the time
leads to unnecessary network performance deterioration. and space complexity for the controller and switches in our
This means that the current routing protocols lack the intelli- proposal. We evaluate the network performance of our
gence to learn from previous experiences. The reason for this proposal in Section VII. Finally, Section VIII concludes
problem is that conventional routing protocols are hardware- the paper.
based, which does not allow the ability of reconfiguration.
And increasing the intelligence in routing protocols also II. RELATED RESEARCH WORK
means the growing computation overhead. On the other The performance of deep learning has been significantly
hand, the controllers in SDCSs manage the networks accord- improved since Hinton et al. [15] proposed the greedy layer-
ing to the installed applications and are equipped with multi- wise training method to pre-train the deep belief architectures.
ple computation platforms, which makes it possible to design As more layers in the structures can represent a more complex
the intelligent routing algorithms. relationship between the input and output, deep learning has
Recently, the deep learning technique has been evaluated become an efficient tool to explore the unknown relationships
to outperform human beings in many areas, such as the board among a number of factors. Besides the academic research on
game go, natural language processing, and so on [13]. In the its applications in image classification and nature language
fields of communication networks, deep learning technique processing [13], various enterprises have adopted deep learn-
has also been studied to improve the performance. Even ing to promote their products and improve their services. For
though the deep learning based algorithms are concerned example, Apple’s ”Siri” utilized this technique to provide the
with more computation compared with conventional strate- best response to customers’ requests [16]. In the field of com-
gies, this technique still has aroused researchers’ interests munication networks, researchers also attempt to consider this
due to its high accuracy rate as well as the flexibility [14]. To technique to address the emerging network challenges. How-
improve the routing performance, in this paper, we propose a ever, not many achievements have been made due to the diffi-
deep learning based routing strategy which can intelligently culty in characterization of the deep learning structure’s input
compute the paths. The proposal consists of two steps: the and output for defining networking problems [14], [17]. Wang
initial phase and running phase. In the initial phase, the con- et al. applied the deep learning technique to find the features of
troller runs the conventional routing protocol while the the traffic flow data [18]. The results showed that their
switches record the traffic trace, which is utilized by the approach works well for protocol identification and anomalous

VOLUME 9, NO.
Authorized 3, JULY-SEPT.
licensed 1555
2021to: Mepco Schlenk Engineering College. Downloaded on October 09,2024 at 05:53:06 UTC from IEEE Xplore. Restrictions apply.
use limited
 Mao et al.: An Intelligent Route Computation Approach Based on Real-Time Deep Learning Strategy for Software Defined Communication Systems

the information to build a global view of the whole network.

Then, via the installed applications, the controller can gener-
ate the forwarding rules for different traffic flows and install
these rules in the corresponding switches. Each rule can be
expressed in the form of hMatch; Actioni, in which Match
consists of some fields, such as the Media Access Control
(MAC) address or Internet Protocol (IP) address of the
source and destination switches. Action gives the correspond-
ing action the switch should take for definite traffic flow [6].
For instance, the Action field can be dropping or modifying
to enable the switch to balance or shape the traffic [9]. The
Action field can also denote the definite interface through
which the traffic flow should be forwarded. Guided by these
rules, when a traffic flow arrives, the switch can analyze the
FIGURE 2. The considered structure of SDCS.
first packet’s header and search the corresponding action in
its forwarding table. If the switch cannot find any entry
protocol detection. In [17], He et al. present an efficient green matching the flow, it will send the header of the first packet
resource allocation algorithm based on the deep reinforcement to the controller. Therefore, the controller will update the
learning, which can achieve high Quality of Experience (QoE) installed rules on the switch.
performance. To meet the fast convergence requirement of
the future backbone network, our earlier work [19] proposed a B. STATE OF THE ART CONVOLUTIONAL NEURAL
tensor based deep learning approach to solving the routing NETWORKS
problem. We considered utilizing a tensor to arrange the Figure 3 depicts the structure of a CNN, which mainly con-
multiple parameters related to routing performance and the sists of three parts: convolutional layers, pooling layers, and
simulation results evaluate the efficiency of the deep learning fully connected layers. For the purpose of classification, a
strategy. However, the performance of the deep learning struc- softmax regression process is usually conducted on the out-
tures in [18], [19] depends on the supervised training which put of the last fully connected layer to get the final out-
require massive labeled training data. In this paper, we con- put [21], [22]. The training process of the CNN can be
sider a real-time training fashion to train the proposed CNNs divided into two value propagation steps: the forward propa-
that only requires a few data before applying the CNNs for gation and the backward propagation. The forward propaga-
routing computation. While the controller utilizes the CNNs tion represents the input information is transferred layer by
to choose the paths, the traffic trace is recorded by the switches layer until an output vector is given, while the backward
to periodically retrain the CNNs to improve the prediction propagation utilizes the output value to adjust the weights
accuracy. and biases of the CNNs. To describe the details more clearly,
in the following paragraphs, we consider two layers labeled
III. PRELIMINARIES as the ðl  1Þth and lth layers as an example to explain every
After introducing the previous related research works, we operation. The input and output of the convolution and pool-
give some preliminary knowledge on SDN and CNNs in this ing operations are usually represented by a three dimensional
section. matrix, while those of the fully connected layers and softmax
operation are denoted as vectors. Moreover, as an activation
A. SOFTWARE DEFINED NETWORKING operation exists in most layers of the CNNs, we adopt U and
Figure 2 shows the structure of an SDCS, of which the data A to denote the values of each layer before and after activa-
plane is a wireless communication network. Different from tion, respectively.
traditional wireless networks which consist of various kinds The convolution operation is to extract the distinguished
of infrastructures like routers, switches, in this network, all features of the input, of which the parameters (weights and
the switches in the data plane can be unified with simple biases) consist of a set of learnable filters. We use a
structures since they just play the role of forwarding packets M  N  P matrix to denote every filter. Every filter is small
according to the installed rules. The complex computation spatially, but extends through the full depth of the input vol-
tasks to generate various rules are outsourced to the SDN ume. During the forward pass, each filter slides across the
controller which holds a global view of the network and width and height of the input volume and a convolution oper-
communicates with the application plane [20]. To clearly ation is conducted between the filter and the area of the input
describe the working procedures of the network, we can take volume covered by the filter. The convolution result of the
the design of forwarding rules as an example. The switches input volume and each filter is named as a feature map. If we
in the data plane conduct a neighbor discovery process and use W ðl1 Þ to denote the filters and the kth filter is represented
ðl Þ
send the controller the obtained information about the neigh- by Wk 1 , the obtained feature map by the convolution opera-
bors and link weights. As a result, the controller can utilize tion can be shown as follows.

1556 VOLUME
Authorized licensed use limited to: Mepco Schlenk Engineering College. Downloaded on October 09,2024 at 05:53:06 UTC from IEEE 9, NO.
Xplore. 3, JULY-SEPT.
Restrictions 2021
apply.
Mao et al.: An Intelligent Route Computation Approach Based on Real-Time Deep Learning Strategy for Software Defined Communication Systems

FIGURE 3. The structure of CNN.

ðl Þ ðl Þ ðl Þ shown in Eqs. (4) and (5). The details of the backward propa-
1
ui;j;k ¼ ðU ðl1 1Þ  Wk 1 Þði; jÞ þ wbk1
gation method can be referred to [23].
P X
X M X
N
ðl1Þ ðl Þ
(1)
¼ wm;n;p aiþm;jþn;p þ wbk1 ; @JðW; BÞ
p¼1 m¼1 n¼1 w :¼ w þ h ; (4)
@w
ðl1 Þ ðl1 Þ
ai;j;k ¼ f ðui;j;k Þ; (2) @JðW; BÞ
b :¼ b þ h ; (5)
ðl1 Þ @b
where f ðÞ is the activation function and ai;j;k
is the activated
value of the unit in the ith row and jth column of the feature
ðl1 Þ ðl Þ where w and b represent any weight and bias, respectively.
map. Therefore, ui;j;k is the value before activation. wbk1
denotes the bias of the kth filter and is usually a single IV. PROBLEM STATEMENT AND MODEL DESIGN
ðl1 1Þ
numeric value. aiþm;jþn;d is the activated value of unit in the With the increasing number of users and rapidly changing net-
ði þ mÞth row and ðj þ nÞth column. The most commonly work environment, global networks are confronted by many
used activation function is the Rectified Linear Unit (ReLU) challenges. To meet the future network requirement, SDN has
function. been regarded as the next generation network paradigm since
After every convolution operation, the output consists of the separation of complex control logic and data forwarding
several feature maps. Then, to progressively reduce the spa- significantly simplifies and unifies the structures of the
tial size in order to reduce the number of parameters and switches. Moreover, the well-defined programmable interface
computation, the operation named downsampling is con- increases the network flexibility. However, current packet for-
ducted through inserting a pooling layer in-between the suc- warding algorithms still follow conventional fixed-rule-based
cessive convolutional layers. The final pooling layer is routing protocols, e.g., minimum or maximum metric values,
connected to the fully connected layer, in which every unit is resulting in the same decision when similar traffic patterns
connected to all units of the previous layer. As the CNN is occur, even though the decision has been proved wrong. To
usually used for classification, if we assume that there are v describe this more clearly, we can take the data plane as shown
kinds of different results, then we can use a vector consisting in Figure 2 as an example. Here, it is reasonable to assume that
of v binary values to represent the result [22]. And in the vec- some switches, S1 , S2 , and S3 , in the data plane, generate pack-
tor, only one unit must have the value of 1, of which the ets destined for the switch S8 . And the central controller choo-
order represents the result. To get the final vector, a softmax ses the best paths which have the minimum hop numbers.
regression process is necessary to be conducted on the final According to the network topology, it is easily understood that
fully connected layer. Since the CNN is trained in a super- the central controller chooses S5 as the next node for source
vised manner, the purpose of the training is to maximize the switches, S1 , S2 , and S3 , to send packets to S8 . Therefore, when
possibility of training data, which can be expressed as a loss three source switches send packets to S8 along with their short-
function as shown in Eq. (3) [22]. est paths concurrently, the joint router of these paths, S5 , easily
! becomes the network bottleneck, leading to the degradation of
1 X m X k
network performance [24]. Even though the congestion can be
JðW; BÞ ¼  1fy ¼ jglog pi ðxi Þ ; (3)
m i¼1 j¼1 alleviated after the switches inform the controller about the
congestion and the controller updates the paths for S1 , S2 , and
where W and B represent the weights and biases of the CNN. S3 , this situation can happen again while similar traffic patterns
m is the number of training data. xi is the ith input training appear. Seriously, when source switches S1 , S2 , and S3 have
data. Here, 1fg is the indicator function and its values satisfy burst traffic patterns as shown in Figure 4(a) and all these pack-
that 1fa true statementg ¼ 1 and 1fa false statementg ¼ 0. ets are destined for S8 , the joint switch S5 becomes congested
The purpose of training is to minimize the loss function and the network performance in terms of the packet loss rate
which can be obtained by adjusting the values of the weights and average packet delay deteriorates repeatedly as shown in
and biases of the CNNs. The process which adopts the final Figure 4(b) and 4(c).
output to adjust the weight and bias is named as backward The reason behind this phenomenon is because of the
propagation and fulfilled through the gradient descent as fixed rule based routing strategy which lacks the ability of

VOLUME 9, NO.
Authorized 3, JULY-SEPT.
licensed 1557
2021to: Mepco Schlenk Engineering College. Downloaded on October 09,2024 at 05:53:06 UTC from IEEE Xplore. Restrictions apply.
use limited
 Mao et al.: An Intelligent Route Computation Approach Based on Real-Time Deep Learning Strategy for Software Defined Communication Systems

FIGURE 5. The input of the CNN in our proposal.

the traffic congestion, we can choose the most direct sign of

the traffic situation, the traffic pattern, as the input. We can
merge the values of different features of the traffic pattern into
a three dimensional matrix as shown in Figure 5. And the three
dimensions represent the router ID, the time interval, and dif-
ferent features of traffic patterns, respectively. If we use TP to
denote the input matrix, then the value of unit, tpijk denotes the
value of switch Si ’s feature k at the jth time interval. The size
of the router ID dimension depends on the number of consid-
ered switches, while that of the time dimension is decided by
many factors, such as the network size, the required accuracy
rate, and so on. The number of chosen time intervals should be
reasonable since traffic patterns of too few time intervals leads
to low prediction accuracy and too many time intervals cause
high computation burden. The number of traffic pattern fea-
tures is concerned with our purpose. As the deep learning
architectures in our proposal are utilized for routing, we choose
the packet generation rates and the switches’ remaining buffer
sizes as our considered features. Considering the input design,
suppose the whole network as an image and different traffic
pattern features as the different color channels of the image.
Therefore, it is reasonable to choose the CNN shown in
Figure 3 as the deep learning structure due to its wide applica-
FIGURE 4. An illustrative example: when switches S1 , S2 , and
tions to process the images. For the output design, since there
S3 choose S5 as the next node to destination S8 , S5 will be the
bottleneck, which means that traffic congestion will easily are several paths for each Origin Destination (OD) pair in our
happen to S5 . proposal, the chosen paths in every round can be regarded as
a paths combination, which consists of one path for each OD
self-reconfiguration. According to the conventional routing pair. Then, each CNN can represent one paths combination
algorithm, the controller always chooses S5 as the next node and the output of the CNN should indicate whether the paths
for S1 , S2 , and S3 since these paths have the minimum hop combination will be congested or not. In our proposal, we can
numbers, despite of the fact that this decision has been use a two dimensional vector to denote the output. We can
proved wrong many times. If the router can adjust its config- set ð1; 0Þ as the notation of congestion and use ð0; 1Þ to denote
uration according to the previous experience, the joint node that the paths combination will not be congested in the next
can be easily avoided. However, conventional proprietary round. Moreover, if the trained CNN outputs results of (0, 0)
hardware based router architecture does not allow the recon- or (1, 1), it means that the training is not effective. Therefore,
figuration of the routing rules without redesigning the hard- we can make some adjustment of the architecture or accumu-
ware architecture. Furthermore, to increase the adaptation of late more training data. To adapt to the changing traffic pat-
the routing strategies is concerned with more complex algo- terns, we consider the online training manner which means the
rithms. On the other hand, the central controller equipped adopted CNNs will be periodically trained with the real-time
with a large computation capacity in the SDCSs can act as traffic patterns. Therefore, the CNNs can adjust its parameters
the platform to run the improved routing algorithms. through the training with new traffic trace.
In this paper, we consider utilizing deep learning to improve
the routing strategy as the deep learning technique is promising V. PROCEDURES OF OUR PROPOSAL
in the future communication network management. To design After introducing our strategy to the problems, in this part,
a deep learning based routing protocol, the first step is to we describe the detailed procedures. Our proposal can be
choose the suitable deep learning structure and define its input divided into two phases: initial phase and running phase. In
and output. Since our goal is to manage multiple paths to avoid the initial phase, we construct a CNN for each paths

1558 VOLUME
Authorized licensed use limited to: Mepco Schlenk Engineering College. Downloaded on October 09,2024 at 05:53:06 UTC from IEEE 9, NO.
Xplore. 3, JULY-SEPT.
Restrictions 2021
apply.
Mao et al.: An Intelligent Route Computation Approach Based on Real-Time Deep Learning Strategy for Software Defined Communication Systems

inaccessible, the neighbor switch can upload the information

to the controller for updating the path.

Algorithm 1. Initial Phase

Input: network topology.
Output: CNNs.
1: The controller generates the global view of the whole
network according to the network topology, compute
the paths combination matrix C with graph theory;
2: for each paths combination ck in C do
3: Controller creates a CNN, CNNk
4: end for
5: for each path update interval tu do
6: The controller computes the best paths and generates
the forwarding rules, and then installs the rules on
corresponding switches
FIGURE 6. The process in the running phase. 7: for each traffic patterns recording interval d do
8: Every switch forwards packets according to the
combination and run the conventional routing protocols to installed rules, records the traffic patterns, and cal-
obtain some data to train these CNNs before utilizing the culates the delay for each received packet.
CNNs to choose the paths combination in the running phase. 9: Every switch calculates the delay of the paths des-
As mentioned previously, the CNNs will be periodically tined for itself, sends the traffic pattern and path
retrained for learning the new network experiences. There- delay values to the controller.
fore, besides the routine path update with CNNs, the running 10: end for
phase consists of two periodically conducted process: data 11: The controller constructs the input traffic patterns of
collection and CNN retraining. Here, it should be noted that the CNN, TP.
12: if the delay for any path of ODij , di;j > threshold
the cycle time of the three process is decided according to
then
the performance. To describe the three process more clearly,
13: y ¼ ð1; 0Þ
we utilize d, tu and tr to denote the cycle time of recording 14: else
traffic patterns, updating paths, and retraining CNNs, respec- 15: y ¼ ð0; 1Þ
tively. And they should satisfy some relationships which can 16: end if
be assumed that tr ¼ n1 tu and tu ¼ n2 d (n1 and n2 are both 17: The controller can generate a set of data ðTP; yÞ for
integers). The detailed procedures are shown in Figure 6, and current paths combination ck
Algorithms. 1, 2 and 3 present the pseudocodes of our pro- 18: Every switch conducts a signaling process, sends the
posal. In the following, we will introduce the two phases link weight to the controller;
according to the figure and algorithms. 19: end for
20: Controller trains all the CNNs with the obtained train-
A. INITIAL PHASE ing data
Since the purpose of initial phase is to obtain some data to
train the CNNs for the considered paths combinations, the Apart from the packet forwarding process, the controller
initial phase consists of two process: utilizing the conven- also needs to construct and train CNNs for choosing the net-
tional routing protocols to forward packets and recording the work paths combinations. The procedures consist of two steps.
traffic trace, and training the CNNs. We explain the two pro- The first step happens at the very beginning when the control-
cess in details one by one. ler first gets the global view of the whole network and the
First, the switches execute a neighbor discovery process weight values of all links as shown in Step 1 in Algorithm 1.
and send the neighbor information to the central controller. With the information, the controller can compute multiple
Then the central controller builds a global view of the whole paths for every OD pair according to some predefined require-
network, including the positions of the switches and their ments, for example, the maximum weight value should not
interfaces. Therefore, as shown in Step 6 in Algorithm 1, the exceed two times of the minimum weight value. Here, we can
controller can run the conventional routing protocols accord- use a vector consisting of n units to save a path in a network
ing to the network topology to choose the best path for every made up of n switches, and utilize a m  n matrix to save the
OD pair in each path update cycle. Then, the controller gen- m paths for an OD pair in a descending priority order defined
erates the packet forwarding rules and installs the rules on by the paths’ metric values. It should be noted that m is the
corresponding switches. Consequently, the switches in the maximum path number among all OD pairs and zeros should
data plane can forward the packets according to the installed be padded in the matrix if the path number is smaller than m
rules. Once some link is congested and some switch becomes for some OD pairs. We can use Pi;j and pki;j to denote the paths

VOLUME 9, NO.
Authorized 3, JULY-SEPT.
licensed 1559
2021to: Mepco Schlenk Engineering College. Downloaded on October 09,2024 at 05:53:06 UTC from IEEE Xplore. Restrictions apply.
use limited
 Mao et al.: An Intelligent Route Computation Approach Based on Real-Time Deep Learning Strategy for Software Defined Communication Systems

matrix for source switch i and destination switch j and the kth destined for itself. Then, all the switches upload the informa-
path, respectively. The controller can choose one path from tion including the traffic patterns and delay values to the con-
each path matrix to construct a paths combination. And all the troller according to Step 9. With these data, the controller can
paths combinations can be saved in a three-dimensional matrix form a matrix TP representing the traffic patterns of all
represented by C in a descending priority order. In this matrix, switches, which will be used as the input of the deep CNNs.
the value of its unit ci;j;k denotes the path order for ODij in the Also, after numerous cycles running the conventional routing
kth paths combination. protocols, the controller can obtain multiple sets of delay val-
ues for each paths combination in C with different traffic pat-
Algorithm 2. Using CNNs to Choose Paths Combination terns. Therefore, for each paths combination, the central
during Each tu controller can judge whether it is congested or not according to
some pre-defined standard, for example, the threshold of the
Input: CNNs
congestion can be two times of the minimum delay value. As
Output: ðx; yÞ (x represents the traffic patterns, TP, y denotes
shown in Steps 12 to 16, if the delay of any chosen path pij
the labels of each paths combination)
1: for p ¼ 1; :::; n1 do exceeds the threshold, it means the chosen paths combination
2: for each paths combination ck do ck is congested, we can get one set of training data for CNNk :
3: Controller conducts a forward propagation process the input is the traffic patterns in previous update interval and
by inputting TP to CNNk and output y. the output is ð1; 0Þ, as we can only use the traffic patterns in
4: if y is (0,1) then last tu to decide the paths in next tu .
5: The paths combination ck is chosen.
6: end if B. RUNNING PHASE
7: break After getting initialized, the CNNs will be applied for routing in
8: end for the running phase to replace the traditional routing protocols.
9: Controller uses the chosen paths combination to gen- Moreover, since we utilize a real-time learning strategy as men-
erate the rules and installs the rules on the corre- tioned in Section IV, the CNNs in our proposal will be periodi-
sponding switches cally retrained with real-time data. Therefore, this phase can
10: for each traffic patterns recording interval d do consist of three parts as shown in Figure 6: data collection, rout-
11: Every switch forwards packets according to the
ing judgement, retraining CNNs, which will be discussed next.
installed rules, records the traffic patterns, and cal-
culates the delay for each received packet.
1) DATA COLLECTION
12: Every switch calculates the delay of the paths des-
Besides forwarding packets all the time, as shown in Steps 10
tined for itself, sends the traffic pattern and path
to 13 in Algorithm 2, every switch in the data plane keeps
delay values to the controller.
collecting the data of traffic patterns in each d as the input of
13: end for
14: The controller constructs the input traffic patterns of CNNs. The switches also calculate and record the delay val-
the CNN, TP. ues when receiving packets. And during each update interval
15: if the delay for any path of ODij , di;j > threshold tu , every switch uploads these data to the central controller,
then and the central controller addresses the data and utilizes for
16: y ¼ ð1; 0Þ two purposes. First, the traffic patterns in the previous path
17: else update interval are adopted as the input of CNNs to choose
18: y ¼ ð0; 1Þ the path for next tu . Second, the controller utilizes the traffic
19: end if patterns and delay values for retraining the CNNs in next tr .
20: The controller can generate a set of data ðTP; yÞ for For example, if the delay of paths combination ck exceeds
current paths combination ck the threshold when the traffic pattern is TP, then the control-
21: end for ler gets one set of data for retraining CNNk , and the input
and output are TP and ð1; 0Þ, respectively.
After obtaining all the paths combinations, the controller
constructs the CNNs as shown from Steps 2 to 4 in Algo- 2) ROUTING JUDGEMENT
rithm 1. Here, we can use CNNk to denote the CNN for paths Since it receives the traffic patterns from all the switches dur-
combination ck . Since the CNNs will be utilized for routing in ing the whole packet forwarding process, the controller can
the running phase, we need to get some data to train our CNNs organize the traffic patterns in the form of CNN’s input as
in the initial phase. As mentioned previously, in every d, each explained in Section IV. Therefore, at the beginning of the
switch records its traffic patterns including the traffic genera- kth update interval, tu , the traffic pattern data of ðk  1Þth
tion rate and the remaining buffer size as shown in Step 8 in update interval are utilized as the input to CNNs to determine
Algorithm 1. To judge whether the path is congested or not, whether the paths combination will lead to congestion or not.
every switch also needs to calculate and record the delay when As the paths combinations are saved in the descending prior-
receiving packets destined for itself. Therefore, in each tu , ity order, the controller will judge these paths combinations
every switch can calculate the delay values for the paths one by one. As shown from Step 2 to Step 8 in Algorithm 2,

1560 VOLUME
Authorized licensed use limited to: Mepco Schlenk Engineering College. Downloaded on October 09,2024 at 05:53:06 UTC from IEEE 9, NO.
Xplore. 3, JULY-SEPT.
Restrictions 2021
apply.
Mao et al.: An Intelligent Route Computation Approach Based on Real-Time Deep Learning Strategy for Software Defined Communication Systems

the judgement process for each paths combination can be ful- In the controller, the deep learning related computations cost
filled by conducting a forward propagation of the corre- most of the resource. The values of computation and storage
sponding CNN with the traffic patterns as the input. The costs depend on the architectures of the CNN, which can be
detailed computation process has been introduced in only decided by trial and error. Therefore, we focus on the
Section III-B. As shown in Steps 4 to 7, if the result of time and space complexity analysis. According to the analysis
CNNk is ð0; 1Þ, which means the paths combination ck will in [25], the computation overhead of training and running one
not be congested, then the controller chooses the paths com- CNN at a time is dependent on the number of nodes in its input
bination for routing in the next tu and the remaining paths layer. Specifically, if n denotes the number of switches in the
combinations will not be considered since they have lower network and each path update interval equals k traffic pattern
priorities. It should be noted that the computation for judging recording intervals, the input layer of each CNN consists of kn
each paths combination is simple and the time cost is negligi- units. Then the time complexity can be denoted as Oðk2 n2 Þ.
ble compared to d. In this paper, we do not consider the delay As k is usually negligible compared with n in our considered
caused by the judgement process. SDCS, the time complexity can be simplified as Oðn2 Þ. The
traditional shortest path strategies, such as the Dijkstra algo-
Algorithm 3. Retraining the CNNs rithm, also have a time complexity not less than Oðn2 Þ. There-
Input: ðx; yÞ (x represents the traffic patterns, TP, y denotes fore, the computation cost of training one CNN with one set of
data is comparable to conventional algorithms. The most com-
the labels of each paths combination)
putation-consuming part of training a deep learning architec-
Output: Updated CNNs
1: for each paths combination ck do ture is that massive data need to be adopted to repeat and
2: Controller trains CNNk with its training data ðx; yÞ iterate the training. However, in our considered proposal, we
3: end for consider the real-time training manner, meaning that each
time, we just use a few sets of accumulated data to retrain sev-
eral CNNs. Compared with the traditional method which trains
all the CNNs one time, the increase of computation cost is still
3) RETRAINING PHASE
limited. This analysis is also applicable to the storage cost.
As mentioned in the previous section, in our proposal, the rout-
Therefore, the space complexity to save the recorded traffic
ing strategy keeps learning from the experiences, which is ful-
patterns and delay values is OðknÞ  OðnÞ.
filled by periodically updating the weight matrices with the
Compared with the training process, to run the deep learning
newly generated traffic trace during the packet forwarding pro-
based proposal, it costs much less computation and storage
cess shown in Algorithm 3. The retraining of the CNNs in the
resource since we just need to utilize one set of traffic pattern
initial phase is nearly the same as that in the initial phase. And
to conduct the forward propagation of several CNNs. And this
compared with the training in this phase, the retraining is based
process has no iteration or repetition. Therefore, the time and
on the previous training, which means that the weights of every
space complexity are just Oðn2 Þ and OðnÞ, respectively.
CNN have reasonable values and the training has less itera-
In our proposal, the switches have the same operations
tions. To more clearly explain the two training process in these
during the training and running periods, meaning the same
two phases, we can think that in the initial phase, the CNNs
computation and storage costs. Moreover, the switches do
are trained to get the basic knowledge about how to choose the
not need to conduct any deep learning related computations
paths combination, while in the running phase, the CNNs
and are just responsible for recording their own traffic pat-
are trained to update and strengthen their knowledge. As the
terns and delay values of the paths from the source nodes to
retraining is a time-consuming process, here, we can assume
themselves. Therefore, we just need to analyze the storage
that the time cost for the retraining process is Dtr . Then, as
complexity, which is Oðk þ nÞ  OðnÞ. Thus, it can be found
shown in Figure 6, before the retraining process is finished, the
that the storage cost for the switches is reasonable.
controller still utilizes the CNNs before retraining to judge the
According to the above analysis, we can find that the deep
paths combinations while the updated CNNs can be adopted
learning technique is related to more computation and storage
once the retraining process is finished.
costs compared with traditional strategy. However, our consid-
ered online training manner can not only increase the self-
VI. COMPLEXITY ANALYSIS adaptation of the CNNs to the traffic changes, but also signifi-
In this section, we make some analysis about the time and cantly alleviate the costs for the controller. Therefore, in the
space complexity of our proposal. As we mentioned earlier, paper, we consider the controller consisting of the CPU and
in our proposal, the switches record their traffic patterns and GPU pools, which can accelerate the computation process.
the delay values of different paths, which are sent to the con-
troller for the training and running of the considered CNNs VII. PERFORMANCE EVALUATION
in the controller. Since all the training and running tasks are This section evaluates our proposal in terms of network perfor-
conducted by the controller, most of the computation and mance through the simulation based on C++ [19]. Since all
storage costs happen in the controller. We first focus on the the computation is conducted on a workstation with Intel Core
controller part and then the switches. i7-6900K CPU, 64GB Random Access Memory (RAM), and

VOLUME 9, NO.
Authorized 3, JULY-SEPT.
licensed 1561
2021to: Mepco Schlenk Engineering College. Downloaded on October 09,2024 at 05:53:06 UTC from IEEE Xplore. Restrictions apply.
use limited
 Mao et al.: An Intelligent Route Computation Approach Based on Real-Time Deep Learning Strategy for Software Defined Communication Systems

TABLE 1. The Parameters of the Considered CNN Structure.

input layer conv1 conv2 fc1 fc2 output layer

width 3 width 3
width 3 height 3 height 3 #node 100 #node 15 #node 2
channel 2 channel 2
stride 1 stride 1
height 10 padding width 1 padding width 1 active relu active relu active softmax
padding height 1 padding height 1
#filter 20 #filter 30
channel 2 active relu active relu initialize xavier initialize xavier initialize xavier
initialize xavier initialize xavier

Nvidia Geforce TitanX GPU, it is reasonable to restrict the packet generation rate and remaining buffer size of each
simulation to a small size network. Therefore, we consider a switch as two channels of the CNN. In each channel, every
scenario of 3  3 wireless heterogeneous network as the data switch records the data in last 10 updating intervals. In
plane and a PC as the central controller which has been shown conv1, we have 20 filters while conv2 has 30 filters, and the
in Figure 2. We consider that the controller manages the size of each filter is 3  3. The padding parameter and the
switches in the form of out of the band. Therefore, independent step size are both 1. The two fully connected layers consist
connections between the central controller and the switches of 100 nodes and 15 nodes, respectively. We consider Xavier
should be established for the transmission of control messages. initialization [28] to set the initial values of all weights and
And the congestion in the data plane does not affect the trans- biases. The accuracy rate of this CNN structure after training
mission of control messages. It is worthwhile to note that this reaches 98.7%, which is sufficient for our proposal.
scale of simulation is sufficient enough to demonstrate that our In the first simulation, we compare the network performance
proposal outperforms conventional routing protocols such as before and after utilizing our deep learning-based proposal.
IS-IS, OSPF, and RIP. In this network, the switches S1 , S2 and The packet generation process in three source switches satis-
S3 generate packets destined for S8 . In order to increase the fies the Possion distribution. And the whole simulation lasts
spectral efficiency, we consider a WLAN system that simulta- about 1,200s while the initial phase and running phase both
neously uses multiple bands such as 2.4GHz and 5GHz [26], share half of the simulation. The average packet generation
[27]. The link bandwidth and the buffer size of each switch are rate is 180Mbps. In the initial phase, the central controller runs
set to 480Mbps and 10MB, respectively. In our simulation, the the conventional routing protocols to generate data for training
sizes of each data packet and signaling packet are 1kb and the CNNs. Then, the CNNs are adopted in the controller to
512b, respectively. The time slot (d) in the simulation is 1s and choose the paths combinations in the running phase. Figure 7
the path updating interval (tu ) consists of only 1 time slot while (a) and 7(b) show the network performance in terms of packet
the retraining time interval (tr ) consists of 100 time slots. loss rate and average packet delay. In the two figures, we can
In our simulation, the structure of CNN after training and find a significant decrease after the application of trained
the parameters have been shown in Table 1. We can find that CNNs into routing, meaning that our proposed CNNs learn to
each CNN consists of 2 convolutional layers (denoted as avoid the congested paths from previous experiences. More-
conv1 and conv2, respectively) and 2 fully connected layers over, the values of packet loss rate and average packet delay
(represented as fc1 and fc2, respectively) as well as the input are still decreasing until reaching the lower bound. This indi-
and output layers. Since the size of the input layer is limited cates that our proposed CNNs are retrained periodically to
because of the considered network size, the pooling layer is learn the new experience, which helps to increase its knowl-
not necessary in our CNNs. In the input layer, we adopt the edge on routing and improve the SDCS performance.

FIGURE 7. The network performance before and after training.

1562 VOLUME
Authorized licensed use limited to: Mepco Schlenk Engineering College. Downloaded on October 09,2024 at 05:53:06 UTC from IEEE 9, NO.
Xplore. 3, JULY-SEPT.
Restrictions 2021
apply.
Mao et al.: An Intelligent Route Computation Approach Based on Real-Time Deep Learning Strategy for Software Defined Communication Systems

FIGURE 8. The network performance comparison between the conventional routing protocol and our proposal in terms of packet loss
rate and average packet delay.

same at t ¼ 300s, which means that the CNNs have

acquired the knowledge on routing after a few times of
training. After that, the accuracy of CNNs in our proposal
gets continuously improved through the periodical retrain-
ing. Therefore, the performance of our protocol outper-
forms the conventional routing protocol in terms of both
the packet loss rate and average packet delay after
t ¼ 300s. Moreover, our proposal keeps improving the
network performance while the performance of OSPF
remains nearly unchanged. This happens because the peri-
odical retraining increases the CNNs’ knowledge for bet-
FIGURE 9. The throughput comparison in the considered SDCS. ter routing while the conventional routing protocol is
based on fixed rules. To further compare our proposal
To compare our proposal with conventional routing with the conventional routing protocol, Figure 9 demon-
protocols, we consider the network traffic patterns gener- strates the network throughput of our proposal and OSPF.
ated by the switches S1 , S2 and S3 are similar to Figure 4 It can be noticed that the network throughput of our pro-
(a). We consider the OSPF algorithm as a benchmark posal is nearly twice than that of OSPF, which can fur-
method. And in the simulation utilizing our proposal, the ther demonstrate the advantages of our proposal over the
initial phase lasts a short time to get a few data for train- conventional routing protocol.
ing the CNNs. To increase the complexity, the start time, In order to further evaluate the performance of our pro-
cycle time, and the amplitudes of the traffic pattern posal under varying network environments, we conduct the
curves are randomly set for the three switches. Figure 8 simulations with the increasing packet generation rate of
(a) and 8(b) compare our proposal and conventional rout- every source switch from 40Mbps to 400Mbps and compare
ing protocols in terms of the packet loss rate and the the packet loss rate, average packet delay, and network
average packet delay. From both results, we can find that throughput of our proposal and conventional routing proto-
the performance of our proposal and OSPF are nearly the cols (OSPF) as shown in Figure 10. It should be noted that

FIGURE 10. Comparison of SDCS performance under different packet generation rates in our proposal and the bencmark methods
(OSPF) in terms of packet loss rate, average packet delay, and throughput.

VOLUME 9, NO.
Authorized 3, JULY-SEPT.
licensed 1563
2021to: Mepco Schlenk Engineering College. Downloaded on October 09,2024 at 05:53:06 UTC from IEEE Xplore. Restrictions apply.
use limited
 Mao et al.: An Intelligent Route Computation Approach Based on Real-Time Deep Learning Strategy for Software Defined Communication Systems

the packet loss rates with the two routing strategies are both 0 REFERENCES
when the packet generation rate is 40Mbps. In Figure 10 [1] J. Liu, S. Zhang, N. Kato, H. Ujikawa, and K. Suzuki, “Device-to-device
(a), it can be clearly found that the network running the communications for enhancing quality of experience in software defined
multi-tier LET-A networks,” IEEE Netw., vol. 29, no. 4, pp. 46–52,
conventional routing algorithm gets congested when the Jul. 2015.
packet generation rate is just above 160Mbps while our [2] H. Guo, J. Liu, Z. M. Fadlullah, and N. Kato, “On minimizing energy con-
proposal can still successfully transfer all the packets sumption in fiwi enhanced LTE-A HetNets,” IEEE Trans. Emerging Top.
Comput., vol. 6, no. 4, pp. 579–591, Oct. 2018.
when the packet generation rate is 160Mbps. When the
[3] L. Qiang, J. Li, and C. Touati, “A user centered multi-objective handoff
packet generation rate is 280Mbps and 400Mbps, the scheme for hybrid 5G environments,” IEEE Trans. Emerging Top. Com-
SDCS using our proposal also gets congested which can put., vol. 5, no. 3, pp. 380–390, Jul.-Sep. 2017.
be explained by the switches’ limited buffer size and link [4] I. T. Haque and N. Abu-Ghazaleh , “Wireless software defined network-
ing: A survey and taxonomy,” IEEE Commun. Surveys Tuts., vol. 18,
bandwidth. On the other hand, the result can still demon- no. 4, pp. 2713–2737, May 2016.
strate that compared with the conventional routing proto- [5] I. F. Akyildiz, A. Lee, P. Wang, M. Luo, and W. Chou, “Research chal-
col, the proposed CNNs can make the better routing lenges for traffic engineering in software defined networks,” IEEE Netw.,
vol. 30, no. 3, pp. 52–58, May 2016.
decision for alleviating the traffic congestion. [6] H. Huang, P. Li, S. Guo, and W. Zhuang, “Software-defined wireless mesh
networks: Architecture and traffic orchestration,” IEEE Netw., vol. 29,
no. 4, pp. 24–30, Jul. 2015.
VIII. CONCLUSION [7] K. Wang, Y. Wang, D. Zeng, and S. Guo, “An SDN-based architecture for
next-generation wireless networks,” IEEE Wireless Commun., vol. 24,
SDN has been viewed as the paradigm of next generation
no. 1, pp. 25–31, Feb. 2017.
network due to its flexibility and conciseness. However, cur- [8] C. Chuang, Y. Yu, A. Pang, H. Tseng, and H. Lin, “Efficient multicast
rent SDN structure mainly utilizes conventional routing pro- delivery for data redundancy minimization over wireless data centers,”
tocols which are based on fixed rules and lacks the IEEE Trans. Emerging Top. Comput., vol. 4, no. 2, pp. 225–241,
Apr.-Jun. 2016.
intelligence to learn from previous experiences. This can [9] D. Kreutz, F. M. V. Ramos, P. E. Verissimo, C. E. Rothenberg, S. Azodol-
lead to the repetition of wrong decisions when similar traffic molky, and S. Uhlig, “Software-defined networking: A comprehensive sur-
patterns happen. The inaccurate path decision results in the vey,” Proc. IEEE, vol. 103, no. 1, pp. 14–76, Jan. 2015.
[10] B. Mao, F. Tang, Z. M. Fadlullah, N. Kato, O. Akashi, T. Inoue, and
network congestion, which leads to further performance dete- K. Mizutani, “A novel non-supervised deep learning based network traffic
rioration. In this paper, we propose a deep learning based control method for software defined wireless networks,” IEEE Wireless
routing strategy which utilizes CNNs to choose the paths Commun. Mag., vol. 25, no. 4, pp. 74–81, Aug. 2018.
[11] H. Hu, H. H. Chen, P. Mueller, R. Q. Hu, and Y. Rui, “Software defined
combinations according to the network traffic trace in an wireless networks (SDWN): Part 1 [guest editorial],” IEEE Commun.
online fashion. This strategy can not only better choose the Mag., vol. 53, no. 11, pp. 108–109, Nov. 2015.
paths combinations according to previous network trace, but [12] N. McKeown, T. Anderson, H. Balakrishnan, G. Parulkar, L. Peterson,
J. Rexford, S. Shenker, and J. Turner, “OpenFlow: Enabling innovation in
also keeps improving its performance through continually campus networks,” ACM SIGCOMM Comput. Commun. Rev., vol. 38,
learning from previous experience. Analysis shows that our no. 2, pp. 69–74, Apr. 2008.
proposal can avoid the congested paths and balance the net- [13] Y. LeCun, Y. Bengio, and G. Hinton, “Deep learning,” Nature, vol. 521,
no. 7553, pp. 436–444, May 2015.
work traffic, resulting in the significant improvement of
[14] N. Kato, Z. M. Fadlullah, B. Mao, F. Tang, O. Akashi, T. Inoue, and
packet loss rate and average packet delay in the SDCS. Thus, K. Mizutani, “The deep learning vision for heterogeneous network traffic
it can be concluded that our proposal outperforms conven- control: Proposal, challenges, and future perspective,” IEEE Wireless
tional routing protocols in SDCSs. Furthermore, with the Commun., vol. 24, no. 3, pp. 146–153, Jun. 2017.
[15] G. E. Hinton, S. Osindero, and Y.-W. Teh, “A fast learning algorithm for
development of the new cellular communication technology, deep belief nets,” Neural Comput., vol. 18, no. 7, pp. 1527–1554,
5G, a growing number of network applications, such as the Jul. 2006.
Internet of Things (IoT) and Vehicle to Everything (V2X), [16] “The ibrain is here and it’s already inside your phone.” [Online]. Available:
https://backchannel.com/an-exclusive-look-at-how-ai-and-machine-learning-
are emerging to provide users with different services. There- work-at-apple-8dbfb131932b\#\.43bf9cm00, Accessed on: Jan. 2018.
fore, the future global network will consist of more heteroge- [17] X. He, K. Wang, H. Huang, T. Miyazaki, Y. Wang, and S. Guo, “Green
neous devices, while the traffic patterns will become more resource allocation based on deep reinforcement learning in content-centric
IoT,” IEEE Trans. Emerging Top. Comput., 2018, doi: 10.1109/
complex. On the other hand, the proposed SDCS neglects TETC.2018.2805718
the heterogeneity of the switches in the data plane, while the [18] Z. Wang, “The applications of deep learning on traffic identification,”
considered realtime learning based routing strategy is self- BlackHat USA, 2015, Available: https://www.blackhat.com/docs/us-15/
materials/us-15-Wang-The-Applications-Of-Deep-Learning-On-Traffic-
adaptable to the network changes. Therefore, our proposal, Identification-wp.pdf
in this paper, can efficiently improve the network perfor- [19] B. Mao, Z. M. Fadlullah, F. Tang, N. Kato, O. Akashi, T. Inoue, and K. Mizu-
mance of the future networks. tani, “A tensor based deep learning technique for intelligent packet routing,” in
Proc. IEEE Global Commun. Conf., Dec. 2017, pp. 1–6.
[20] J. L. Chen, Y. W. Ma, H. Y. Kuo, C. S. Yang, and W. C. Hung, “Software-
ACKNOWLEDGMENTS defined network virtualization platform for enterprise network resource
management,” IEEE Trans. Emerging Top. Comput., vol. 4, no. 2,
The research results have been achieved by ”Research and pp. 179–186, Apr.-Jun. 2016.
Development on Intellectual ICT System for Disaster [21] “CS231n convolutional neural networks for visual recognition.”
[Online]. Available: http://cs231n.github.io/convolutional-networks/,
Response and Recovery”, the Commissioned Research of Accessed on: Jan. 2018.
National Institute of Information and Communication Tech- [22] “Convolutional neural network.” [Online]. Available: http://ufldl.stanford.edu/
nology (NICT), Japan. tutorial/supervised/ConvolutionalNeuralNetwork/, Accessed on: Jan. 2018.

1564 VOLUME
Authorized licensed use limited to: Mepco Schlenk Engineering College. Downloaded on October 09,2024 at 05:53:06 UTC from IEEE 9, NO.
Xplore. 3, JULY-SEPT.
Restrictions 2021
apply.
Mao et al.: An Intelligent Route Computation Approach Based on Real-Time Deep Learning Strategy for Software Defined Communication Systems

[23] J. Bouvrie, “Notes on convolutional neural networks,” 2006, [Online]. ZUBAIR MD. FADLULLAH (M’11-SM’13)
Available: http://cogprints.org/5869/1/cnn_tutorial.pdf, Accessed on: Jan. received the BSc degree with honors in computer
2018. science and information technology from Islamic
[24] Y. Xiao, K. Thulasiraman, G. Xue, and M. Yadav, “QoS routing under University of Technology (IUT), Bangladesh, in
multiple additive constraints: A generalization of the LARAC algorithm,” 2003, and the MSc and PhD degrees in applied
IEEE Trans. Emerging Top. Comput., vol. 4, no. 2, pp. 242–251, information science from Tohoku University, in
Apr.-Jun. 2016. 2008 and 2011, respectively. He is currently an
[25] B. Mao, Z. M. Fadlullah, F. Tang, N. Kato, O. Akashi, T. Inoue, and associate professor with the Graduate School of
K. Mizutani, “Routing or computing? the paradigm shift towards intelli- Information Sciences, Tohoku University, Japan.
gent computer network packet transmission based on deep learning,”IEEE His research interests include 5G, smart grid, net-
Trans. Comput., vol. 66, no. 11, pp. 1946–1960, 1 Nov. 2017. work security, intrusion detection, game theory,
[26] N. Egashira, K. Yano, S. Tsukamoto, J. Webber, and T. Kumagai, “Low quality of security service provisioning mechanism and deep learning. He
latency relay processing scheme for WLAN systems employing multiband received the Dean’s Award and the President’s Award from Tohoku Univer-
simultaneous transmission,” in Proc. IEEE Wireless Commun. Netw. sity in 2011, the IEEE Asia Pacific Outstanding Researcher Award in 2015,
Conf., Mar. 2017, pp. 1–6. and the NEC Foundation Prize for research contributions in 2016. He also
[27] Z. M. Fadlullah, Y. Kawamoto, H. Nishiyama, N. Kato, N. Egashira, received several best paper awards in the Globecom, IC-NIDC, and IWCMC
K. Yano, and T. Kumagai, “Multi-hop wireless transmission in multi-band conferences. He is a senior member of the IEEE.
WLAN systems: Proposal and future perspective,” IEEE Wireless Com-
mun., vol. 26, no. 1, pp. 108–113, Feb. 2019, doi: 10.1109/
MWC.2017.1700148. NEI KATO (F’13) is a full professor and the director
[28] “Understanding xavier initialization in deep neural networks.” [Online]. of Research Organization of Electrical Communica-
Available: https://prateekvjoshi.com/2016/03/29/understanding-xavier- tion (ROEC), Tohoku University, Japan. He has been
initialization-in-deep-neural-networks/, Accessed on: Jan. 2018. engaged in research on computer networking, wire-
less mobile communications, satellite communica-
tions, ad hoc & sensor & mesh networks, smart
BOMIN MAO (S’15) received the BSc degree in grid, IoT, Big Data, and pattern recognition. He has
telecommunications engineering and the MS published more than 400 papers in prestigious
degree in electronics and telecommunications engi- peer-reviewed journals and conferences. He is the
neering from Xidian University, China, in 2012 vice-president (member & global activities) of IEEE
and 2015, respectively. Currently, he is working Communications Society (2018-2019), an editor-
toward the PhD degree in the Graduate School of in-chief of IEEE Network Magazine (2015-2017), an editor-in-chief of the
Information Sciences (GSIS), Tohoku University, IEEE Transactions on Vehicular Technology (2017-), an associate editor-
Japan. His research interests include wireless net- in-chief of the IEEE Internet of Things Journal (2013-), and the chair of IEEE
works, software defined networking, quality of ser- Communications Society Sendai Chapter. He served as a member-at-large on
vice, particularly with applications of machine the Board of Governors, IEEE Communications Society (2014-2016), a vice
intelligence and deep learning. He received the best chair of Fellow Committee of IEEE Computer Society (2016), a member of
paper awards in Globecom’17. He is a student IEEE Computer Society Award Committee (2015-2016) and IEEE Communi-
member of the IEEE. cations Society Award Committee (2015-2017). He has also served as the chair
of Satellite and Space Communications Technical Committee (2010-2012) and
Ad Hoc & Sensor Networks Technical Committee(2014-2015) of IEEE Com-
FENGXIAO TANG (S’15) received the BE degree munications Society. His awards include Minoru Ishida Foundation Research
in measurement and control technology and instru- Encouragement Prize (2003), Distinguished Contributions to Satellite Commu-
ment from the Wuhan University of Technology, nications Award from the IEEE Communications Society, Satellite and Space
Wuhan, China, in 2012 and the MS degree in soft- Communications Technical Committee (2005), the FUNAI information
ware engineering from the Central South Univer- Science Award (2007), the TELCOM System Technology Award from Foun-
sity, Changsha, China, in 2015. Currently, he is dation for Electrical Communications Diffusion (2008), the IEICE Network
working toward the PhD degree in the GSIS, System Research Award (2009), the IEICE Satellite Communications Research
Tohoku University, Japan. His research interests Award (2011), the KDDI Foundation Excellent Research Award (2012), IEICE
include unmanned aerial vehicles system, game Communications Society Distinguished Service Award (2012), IEICE Commu-
theory optimization, and machine learning algo- nications Society Best Paper Award (2012), Distinguished Contributions to
rithm. He is a student member of the IEEE. Disaster-resilient Networks R&D Award from Ministry of Internal Affairs and
Communications, Japan (2014), Outstanding Service and Leadership Recogni-
tion Award 2016 from IEEE Communications Society Ad Hoc & Sensor
Networks Technical Committee, Radio Achievements Award from Ministry of
Internal Affairs and Communications, Japan (2016) and Best Paper Awards
from IEEE ICC/GLOBECOM/WCNC/VTC. He is a distinguished lecturer of
IEEE Communications Society and Vehicular Technology Society. He is also a
fellow of the IEEE and IEICE.

VOLUME 9, NO.
Authorized 3, JULY-SEPT.
licensed 1565
2021to: Mepco Schlenk Engineering College. Downloaded on October 09,2024 at 05:53:06 UTC from IEEE Xplore. Restrictions apply.
use limited