Learning-based Orchestrator for Intelligent Software-defined Networking Controllers

International Journal of Software Engineering & Applications (IJSEA), Vol.11, No.6, November 2020
DOI: 10.5121/ijsea.2020.11602 17
LEARNING-BASED ORCHESTRATOR FOR
INTELLIGENT SOFTWARE-DEFINED
NETWORKING CONTROLLERS
Imene Elloumi Zitouna
University of Tunis El Manar, École Nationale d'Ingénieurs de Tunis, Laboratoire
Systems de Communications, Tunisia
ABSTRACT
This paper presents an overview of our learning-based orchestrator for intelligent Open vSwitch that we
present this using Machine Learning in Software-Defined Networking technology. The first task consists of
extracting relevant information from the Data flow generated from a SDN and using them to learn, to
predict and to accurately identify the optimal destination OVS using Reinforcement Learning and Q-
Learning Algorithm. The second task consists to select this using our hybrid orchestrator the optimal
Intelligent SDN controllers with Supervised Learning. Therefore, we propose as a solution using Intelligent
Software-Defined Networking controllers (SDN) frameworks, OpenFlow deployments and a new intelligent
hybrid Orchestration for multi SDN controllers. After that, we feeded these feature to a Convolutional
Neural Network model to separate the classes that we’re working on. The result was very promising the
model achieved an accuracy of 72.7% on a database of 16 classes. In any case, this paper sheds light to
researchers looking for the trade-offs between SDN performance and IA customization
KEYWORDS
Open vSwitch OVS, Artificial Intelligence, Machine Learning, Supervided Learning, Reinforcement
Learning, Hybrid Orckestrator, Openflow, QoS, QoE, Real time, User Behavior, User Engagements,
Intelligent Software-Defined Networking ISDN.
1. INTRODUCTION
With the development of Software-Defined Anything, software is already penetrating into the
Internet and even controlling the network. For example, a software developer can make a
service/application at the application layer, and directly obtain the original data of the network
device in the physical layer using southbound APIs [1]. Software technologies are already
changing traditional network services that are typically provided by servers via application layer
protocols into new network services that are provided from the network infrastructure (e.g.,
firewall service, QoS service). The network infrastructure also can provide services through open
APIs, allowing network services to benefit operationally by enabling automated provisioning of
network applications with different characteristics. This also helps to ensure that specific
applications are getting the proper network resources that are dynamically allocated to meet
service requirements (QoS, QoE, encryption, etc.) [1]. Due to these advances, many new service
models are emerging in the field of networking (SDN/NFV/SD-WAN).
Artificial Intelligence (AI) and Machine Learning (ML) approaches have emerged in the
networking domain with great promise. They can be clustered into AI/ML techniques for network
engineering and management, network design for AI/ML applications, and system aspects.
AI/ML techniques for network management, operations and automation improve the way we

18
address networking today [10] [11] [12]. They support efficient, rapid, and trustworthy
management operations. Machine learning (ML), is a key feature of future networks mainly the
SDN paradigm.
Our work and interest in softwarization and network programmability fuels the need for
improved network automation, including edge and fog environments. Moreover, network design
and optimization for AI/ML applications address the complementary topic of supporting AI/ML-
based systems through novel networking techniques. Including new architectures and
performance models with great predictive ability in order to guarantee the continuity of QoS/QoE
[2] [3].
In this paper, we elaborate an alternate method and enhancements provided. For the purpose of
tackling the current difficulties, we propose an innovative solution based on the application of
Artificial Intelligence and supervised and reinforcement Machine Learning techniques in SDN.
Through SDN network virtualization, its layered architecture and its programmable interfaces,
overall network control is therefore ensured by the logical centralization of the control function in
the SDN controllers. Network flows are thus controlled using the OpenFlow protocol. In this
work we present, a new reinforcement learning model in an SDN infrastructure. Which adds
"learning" to the control function for better intelligent decision-making at the level of SDN
controllers [7] [8].
These results will ultimately bring us a step closer to overcoming the problem of over
exploitation of network resources, and how they can be deployed more efficiently.
One of the challenges that every machine learning algorithm is faced with is scalability and
validity to large datasets. Our research devoted to applying, hybrid machine learning, which use
an orchestrator and a set of intelligent controllers SDN combine supervised and reinforcement
learning techniques, can improve the key performance metric as well as QoS of different real
time applications and the QoE. Furthermore, prediction methods for behavior and engagements
user combining classification techniques have the potential for creating more accurate results
than the individual methods, particularly for large datasets.
The remainder of this paper is structured as follow: section II discusses a challenge, section III
ISDN and orchestrator quality requirements, section IV describes our approach, section V
presents design of reinforcement learning framework of ISDN, section VI describes the
experimental results of supervised learning of the hybrid orchestrator, testbed, model context,
dataset, data preparation, model selection and training, model creation and fine-tuning. Finally,
section VII conclude the paper.
2. CHALLENGE
Adoption of the Service-Oriented Architecture principle in networking has enabled the Network-
as-a-Service paradigm that is expected to play a crucial role. Network services significantly
impact the performance of higher layer services/applications. The behavior of an end-to-end
network service is the result of the combination of the individual network function behaviors as
well as the behaviors of the network infrastructure composition mechanism.
However, this emerging network services are usually are usually compositions of multiple service
components, APIs or network functions, based on new mechanisms, such as service mesh,
internet of services, and service function chaining, running on the network layer and/or
application layer. Meanwhile, machine learning (ML) has seen great success in solving problems

19
from various domains. In order to efficiently organize, manage, maintain and optimize
networking systems, more intelligence needs to be deployed.
In this case, can we apply the machine learning algorithms in the realm of SDN, from the
perspective of traffic classification, routing optimization, quality of service/quality of experience
prediction and resource management ? How can SDN/ML brings us new chances to provide
intelligence inside the networks ?
The logically centralized control, global view of the network, software-based traffic analysis, and
dynamic updating of forwarding rules, of SDN make it easier to apply our solution of machine
learning techniques in the SDN: "The Intelligent Software defined networking, ISDN".
It is believed that ML has high potential the aforementioned challenges in SDN-based networks,
dynamic service provisioning and adaptive traffic control, as ML is a technology that can
effectively extract the knowledge from data, and then accurately predict future resource
requirements of each virtualized software-based appliance and future service demands of each
user. Therefore, our research tends towards the proposal of a new hybrid orchestrator which
learns the optimal ISDN controller in a multi-ISDN environment.
In another context, in the absence of a mechanism for evaluating the best SDN controller and best
OVS in terms of response time in network congestion situations, centralization at the SDN
controller level operated by OpenFlow increases these response times. Consequently, this will
allow any attacker who has gained access to the administration network to act at any time on the
configuration of the OVS.
It is also conceivable for an attacker to place himself in a man-in-the-middle in order to filter the
commands of a legitimate administrator as well as the feedback of information from the OVS.
Finally, listening to unencrypted or encrypted OpenFlow messages passively provides a great
deal of information about the networks maintained by the controller.
Since the memory resources of the controller are likely to be greater than those of the OVS. This
attack model presupposes the possibility of recording flows on the controller. There is a higher
risk compared to the system master-slaves.
Therefore, without an intelligent prediction mechanism, it is simple for an opposing SDN
controller to usurp the role of master, or to prevent a legitimate SDN controller who wishes to be
master can configure the OVS, or at least have processing times that increase with the traffic rate.
On the other hand, we can measure even longer processing times for large data flows. The goal of
our work is to be able to integrate intelligence into SDN controllers through the application of
machine learning algorithms in terms of routing optimization, traffic classification and resource
management.
3. ISDN AND ORCHESTRATOR QUALITY REQUIREMENTS
The QoS refers to technologies that manages network traffic in order to reduce packet loss,
latency and jitter. They control and manage network resources by defining priorities for specific
forms of information on the network. The QoE is a measurement that determines user behavior
and user engagements and how satisfied the end user is network services [15]. Contrary to QoS,
QoE takes considers the end-to-end connection and applications presently running over that
network connection and how multimedia elements are meeting the end user’s demands.

20
• QoS metrics of OVS Flow table [2] [3] [4]:
- PacketIn: this event is raised whenever the controller receives an OpenFlow packet-in
message from the OVS. This indicates that a packet has arrived at the OVS port and the
OVS either does not know how to handle this packet or the matching rule action implies
sending the packet to the controller. This is where the controller performs the overhead
calculation by incrementing the overhead when the OVS forwards a packet.
- UpStream: this event is triggered in response to the establishment of a new control
channel with a OVS.
- DownStream: this event is fired when a connection to a OVS is terminated, either
explicitly or because of OVS rebooting. We utilize this event handler to output the final
value of measured control overhead.
- TCAM: the OVS support the wire-rate access control list (ACL) and QoS feature with
use of the ternary content addressable memory (TCAM). The enablement of ACLs and
policies does not decrease the switching or routing performance of the OVS as long as the
ACLs are fully loaded in the TCAM. If the TCAM is exhausted, the packets may be
forwarded via the CPU path, which can decrease performance for those packets. Therefore,
our learning can be done with the (QoS and ACL) TCAM percentage of the switches. This
value can be used to determine the reward of the learning agent for the OVS 1.
- Throughput: we aim to measure the maximum flow setup rate a controller can handle per
period of time. Being aware of the amount of packet-in a controller can process ease the
choice of SDN controllers appropriate to master network control charge.
- Latency: represents the time consumed by a controller to process and reply an
outstanding flow request from the OVS.
- Scalability: represents the capacity of a controller to handle a large network with an
increased number of connected hosts without degrading throughput and latency
performance.
- ICMP Round Trip Time (RTT): it is important to evaluate the influence of additional
delay due to communications between OVS and SDN controller for the first ICMP Echo
packet we target the selection of a controller that shows the lower delay.
- TCP and UDP measurements: the purpose of TCP and UDP measurements is to estimate
the impact of delay introduced by the communication between controller and OVS on the
TCP transfer time and UDP packet losses. TCP and UDP evaluation are performed over
multiple network topologies by Mininet. Single, linear, tree and data center network (DCN)
topologies.
• User behavior [5] [6] [9]: is the tracking, collecting and assessing of user data and activities
using monitoring systems. Our purpuse is to determine a baseline of normal activities
specific to the organization and its individual users. They can also be used to identify
deviations from normal. We propose to uses machine learning algorithms to assess these
deviations of user behavior in near-real time.

21
• User engagements [5] [6]: is a process comprised of four distinct stages: 1) point of
engagement, 2) period of sustained engagement, 3) disengagement, and 4) reengagement.
The period of Engagement Attributes are: Aesthetic and Sensory Appeal, Attention,
Awareness, Control, Interactivity, Novelty, Challenge, Feedback, Interest and Positive
Affect. Also we propose to uses machine learning algorithms for specification the new
SLA of user engagement
4. APPROACH OF LEARNING ORCHESTRATOR
Our intention would be to design an SDN architecture composed of multiple controllers where
the choice of the suitable and optimal controller and OVS to be implemented in a specific domain
depends on the controller performance and efficiency to respond to QoS and QoE requirement
imposed by the service application with machine learning environment.
We extract relevant information from the Data flow generated from a SDN (PacketIn, UpStream,
DownStream, TCAM, etc.) and using them to learning, to predicting and to accurately identify
the optimal destination OVS using Reinforcement Learning. Also we extract relevant information
from the Data flow generated from the optimal Intelligent Software-Defined Networking
controllers (CPU Throughput, CPU Latency, Scalability, ICMP Round Trip Time (RTT), TCP
and UDP measurements, etc) with Supervised Learning run in the Hybrid Orchestrator (Figure 1).
Our performance evaluation considers two well-known centralized controllers and multi OVS.
The evaluation in terms of throughput, latency, and scalability is performed with Cbench. With
Mininet and Iperf we evaluate TCP, UDP and ICMP flow over several network topologies. The
main purpose is to predict and to select the ISDN controller that exhibits the highest throughput,
scalability, lowest latency, delay, packet loss, TCP transfer time and most rich in features and
also to predict the optimal OVS with lowest TCAM percentages. This value can be used to
determine the reward of the learning agent for the session.
We extend this hybrid orchestrator by adding the code to install the selected forwarding North
Rules (R1(RRD), R2(SRD), R3(RTPD), etc) created by the learning agent prior to handling any
outcoming traffic.
At the beginning of each emulation experiment, traffic flows are generated into the network
topology, the state of the flow table in the OVS, and the state of the ISDN controllers is observed.
To test the controller’s performance in terms of latency, OVS sends an asynchronous message
toward controller and waits for a reply. Hence, the number of acknowledgments gathered in a test
period of time is used to calculate the average latency. For the throughput measurements, each
OVS transfers a burst of packet-in without waiting for a response to estimate maximum flow
requests a controller can handle per second. Emulated hosts have access to Linux file system
commands. Like “Iperf”. Therefore, network design can easily move to the real system with
minor deployment [5].

22
Figure 1. Hybrid Learning Orchestrator for multi-ISDN proposed
5. DESIGN OF REINFORCEMENT LEARNING FRAMEWORK OF ISDN
In our approach, the configuration of rules can be defined by two decisive parameters, flow
match frequency and flow recentness. Therefore, the RL algorithm can be modeled to obtain the
network configuration that minimizes the long-term control plane overhead [7]. This approach is
modeled using a Markov Decision Process (MDP), which can be represented as the tuple <S, A,
P, R>, where S = {s1, s2, . . . , si } is the state-space, A = {a1, a2, . . . , aj } is the action-space, P
defines the transition probability from state i to state j, and R defines the reward associated with
various actions a ∈ A. The goal of this MDP is to develop a policy π : S → A that maximizes the
cumulative rewards obtained in the long term [4][6]. In the problem that we address in this
research, we can know all the forwarding rules that need to be processed for all tra c ows
transported on the network. For our problem, the state-space, action-space, and reward function
are de ned as follows:
In our approach, we utilize the following RL algorithms:
5.1. Reinforcement Learning Framework of ISDN.
In our design, the Q-Learning algorithm is utilized in our emulation experiments. The decision to
utilize the QLearning algorithm is motivated by the fact that it does not need a model for the
environment, and it updates the Q-values at each time step based on the estimation of action-
value function. This iterative learning process is conducted in discrete time steps in which the
agent interacts with the environment. At each step, the agent interacts with the environment. At
each step, the agent selects an action at ∈ A in the given state st.The agent makes its own

23
decisions to choose an action based on the action selection policy with the goal of maximizing
the expected reward. The notation Q(s, a) denotes the average Q-value of an action a at state s.
While the immediate reward is collected, Q(s, a) can be further refined as:
Q(st,at)= r + Q(st,at) - εQ(st,at)+ γ maxQ(st+1,at+1)
where ε (0 < ε ≤ 1) is the learning rate that determines to what degree the newly obtained
information overrides the old one and γ (0 ≤ γ ≤ 1) is the reward factor that defines the
importance of future rewards and guarantees convergence of the accumulated reward. In our
experiments, we use a discount factor of 0.85 as these values are commonly used in practice.
ALGORITHM : Q-Learning Algorithm
Preconditions:
Initialize Q-table random
TCAM percentage Initialize
state st
Initialize goal 
Procedure:
01: improvement measures = 0
02: repeat
03: for (step = 0; step < learning_iteration;
step++) 04: Get action at from st using 
05: Get parameter from st using 
06:  =  – *(step / learning_iteration)
07: Take action at on the parameter and receive reward r, control overhead C
08: Sample new state st +1 after applied action at
09: Update Qt←r+ Qt −Qt +γ∗maxQt+1
10: Update the corresponding parameter of Qt
11: improvement measures = get_improvement measures
(bestt ++, bestt - - )12: end for
13: until improvement measures > 
The Q-Learning algorithm can have a chance to explore the action space in the process of
searching for a better action that produces a better reward. β is used as the action selection policy
in which the action with the highest Q- value is always picked while the probability of picking
some other action at random is small.
If the action is chosen to be performed due to its Q-value, the associated parameter value will be
updated as well. ε represents the terminal objective of the algorithm. Where this algorithm keeps
track of the chosen parameter corresponding to the Q-value of each action.
This value interval can be divided into four different states as follows:
- State 1: If the value of the TCAM belongs to [0..30]: it is a weak state
- State 2: If the value of the TCAM belongs to [30..50]: it is an ideal and appreciated state.
- State 3: If the value of the TCAM belongs to [50..80]: it is a critical state.
- State 4: If the value of the TCAM belongs to [80..90]: it is a mediocre state
To facilitate the understanding of the different actions of the agent, we can first represent the
states as follows:

24
Table 1. The different states available for each controller
State 1 State 2
State 4 State 4
The agent's actions are modelled as follows:
Action 1 (go up): transition from state 4 to state 1 or from state 3 to state 2 Action 1 (go up):
transition from state 4 to state 1 or from state 3 to state 2 Action 3 (turn right): transition from
state 1 to state 2 or from state 4 to state 3 Action 4 (turn left): transition from state 2 to state 1 or
from state 3 to state 4
6. EXPERIMENTAL RESULTS
6.1. Testbed Setup
The testbed consists of a single machine with (Intel Core i5-7200U CPU @ 3.1 GHz), 8 GB of
memory available. The system runs Ubuntu 16.04 LTS-64 bit with VMware Workstation Pro
version 14.1.1 installed. Controllers, Cbench and Mininet are installed in separate VMs the
allocated memory is respectively (4GB, 2GB, 2GB).
When it comes to Machine learning, there are four main steps that need to be followed which are:
1. Looking at the big picture
2. Getting The Data
3. Preparing the Data for Machine Learning algorithms
4. Selecting and training, and fine-tuning a model The machine learning libraries used are :
1. TensorFlow : is well known for its flexible architecture which allows for the deployment of
computation across a variety of platforms like CPUs or GPUs on a desktop, a server, or a
mobile device.
2. Keras : is a python-based Deep Learning. It works in a different way than other Deep
Learning frameworks. Keras does not support low-level computation like TensorFlow
which is why it uses a tool named Backend.
3. Scikit-learn : is a Machine Learning Library. It includes a wide array of classification,
regression, and clustering algorithms. It was conceived in a way to interoperate with other
mathematical python libraries like Scipy, and NumPy.
4. Jupyter Notebook : is a Web application allowing the creation and sharing of doc- uments
containg live code, visualizations, and narrative text.
5. Anaconda : is a Python and R programming languages distribution which conveniently
installs them additionally to other packages commonly used in Data science and scientific
computing.
We have succeded in training a model which is to be implemented on an SDN Controller. its task
is intelligent control and routing. We will go through these steps one by one while explicitly
explaining the choices we have made throughout the development process [13] [14].
6.2. Model Context
For this model, we used Mininet, a network emulator to generate a Network OVS, controllers,
hosts, and links. in figure .2. The OVS support Openflow protocol for high flexibility when it
comes to customized routing and SDN. Figure 3, 4 and 5 presents the iteration of our algorithm.

25
6.3. Q-Learning Algoritm Outputs
Starting from the database containing the ternary memory consumption rate of a OVS as a
function of time, we applied the Q-learning algorithm and we have it trained several times so that
it converges towards an optimal result and informs us the best OVS in terms of QoS. At each
iteration, the agent stores the corresponding values (Q-Values) resulting from Bellman's equation
in a dynamic array named Q-table. Since all of the study and testing is done on a database of
seven OVS, such a procedure allows us to get two OVS in order to decide which of these two is
the most efficient.
The different rates of memory consumption existing in our database are quite large. The figure. 2
below shows part of the database imported into the working environment.
The three previous cases show us different iterations of the algorithm, which ends up deciding in
the Xth iteration, with X is the length of the database, that the OVS R2 is optimal and we can also
retain it.
6.4. The Dataset
In order to train a model, we need a Dataset. In our case, we used the emulated network on
Mininet to generated Data flows and recuperate them using Wireshark, a free opensource packet
sniffer. A sample of the recuperated data is shown in Table II.
Figure 2. Percentage of ternary memory consumption as a function of time
Figure 3. 1st iteration of the algorithm

26
Figure 4. 5th iteration of the algorithm
Figure 5. Xth iteration of the algorithm
6.5. Data preparation
To prepare our Data, we first start by eliminating useless information from it. As is shown in the
Table II., our Dataset contains many columns that our model’s has no use for, which is the
descriptive information in the last column. Another constraint when it comes to exploiting
datasets is that Machine Learning algorithms can’t handle String input, for that we had to
transform Source IP address, Destination IP address and Protocol columns into integers. This
process can be easily done via simple Python functions. The following Table.I illustrates the
prepared Dataset.
Following this step, we proceed to dividing our Data into two parts: a Training set and a
validation set.
Table 2. A Glimpse of the Raw Dataset

27
Table 3. A Glimpse of the Processed Dataset
6.6. Model Selection and Training
As our Data is labeled, it is natural to choose a Supervised Learning algorithm, all the more so
due to Supervised learning being customizable and frequently used in problems relating to self-
healing networks. This is particularly the case for the OpenDayLight project, a modular open
platform for the customization and the automatization of networks of all sizes. It is a Project that
took birth out the SDN movement, clear focusing on Network programmability.
In the following section, we introduce the libraries used from realization of our model.
6.7. Model Creation and Fine-tuning
When it comes to Creating and Fine-tuning Convolutional Neural Network models comes down
to optimizing the Hyperparameters of the model (e.g. Number de filters, size of each du filter,
activation function of each layer, etc.). There are many ways to do this. One can try to fiddle with
them manually until a great combination is found, which can be quite time-consuming, or he can
try to use pre-installed Scikit-learn search methods like the Randomized search.
In our case, and considering the fact that our Data is not complexe, we used the simplest fine-
tuning method which is Grid search. the resulting optimized CNN structure had the following
architecture in figure .6.
The following is a break-out of the models layers:
• The Input Layer: this layer contains our input values which we want to predict and train the
model with, where each neuron in that layer contains a value took from the sample. In our
case, our dataset contains 6 columns, thus the 6 neurons at the top layer.
• The Output Layer: is the layer will give us an information about the sample we gave to our
neural network. In our case, it would be predicting the IP destination address. As our Data
flow simulation engaged 16 IP addresses, the number of neurons in this layer is 16.
• The Hidden Layer: The hidden layers can be composed of one or many sublayers whose
numbers and hyperparameters where determined by the Grid Search. In our case it contains
two layers.

28
Figure 6. Model Architecture
At this point, we have built our model and pre-processed our Data. All that is left is to compile
the model, train it, and save it using the previously defined functions compile(), model.fit(), and
model.save() respectively.
The evaluation of our model using the validation Data gives an accuracy ratio of 72.7%.
6.8. Problems and limitations
When it comes to Machine Learning, the main tasks are the selection of the algorithm and the
Data, and latter is what we encountered.
• Insufficient Quantity of Data is a Major problem when it comes to training a model. Even
for very simple problems like ours, we need thousands of examples. For more complex
problems like sound recognition or image classification, we would need millions of
examples [16].
• Irrelevant Features: As they say: garbage in, garbage out. Our model will only learn
correctly if the training Data contains many relevant features and not too many
insignificant ones. These problems, more often than not, lead data scientists to over-fitting
their Data to the training set, Optimizing the hyperparameters in search for a better
accuracy ratio at the training without taking into account the global, dynamic context in
which the mode l will be deployed, which usually leads to bad overall performance after
deployment.
CONCLUSIONS AND FUTURE WORKS
Supervised machine learning methods, such as neural network (NN), convolutional neural
network (CNN), and recurrent neural network (RNN) can apply to prediction and classification.
Furthermore, reinforcement learning methods, are tools for generative networks and
discriminative networks.

29
These methods contribute substantially to improving prediction and classification in relevant
applications, but there remain issues and limitations that require further attention from the
research community.
The testing process of Machine Learning on SDN controllers has long been considered critical
for ISPs around the world. Fortunately, our trained-model confirmed the viability and the
potential of the combination of the two aforementioned technologies.
For our case, we were implementing a supervised learning algorithm for intelligent routing and a
reinforcement learning algorithm for to choice the optimal controller. This hybrid learning
solution is being improved constantly as it self adjusts itself to all the customers.
In this paper, we trained a supervised-learning CNN to predictively routing data frames while
arguing the choices taken in the process. We additionally presented the used frameworks
followed by the experimental results. We expend the New Supervised Machine Learning
Methods to the business layer and try to implemented it in an hybrid orchestrator which will have
to efficiently choose which controllers ISDN to use in real-time to maximize performance
indicators with Novel Reinforcement Learning Method where an agent has to take real-time
decisions which is true for our problem. Despite the fact that the machine learning-based
approach performs well in our experiments, further development is needed to expand the
capabilities of our emulation scope. The rest of this work will be the application of Q-learning
algorithm that we proposed. Currently, implementing the user behavior and the user engagements
in the learning labels and model training more thoroughly is the focus of our ongoing research.
REFERENCES
[1] J. Wagner. Why Performance Matters. Accessed: Mar. 22, 2019. [Online]. Available:
https://developers.google.com/web/fundamentals/performance/ why-performance-matters/
[2] M. Seufert, P. Casas, N. Wehner, L. Gang, and K. Li, “Streambased machine learning for real-time
QoE analysis of encrypted video streaming traffic,” in Proc. 22nd Conf. Innov. Clouds, Internet
Netw.Workshops, Feb. 2019, pp. 76–81.
[3] V. Krishnamoorthi, N. Carlsson, E. Halepovic, and E. Petajan, “BUFFEST: Predicting buffer
conditions and real-time requirements of HTTP (S) adaptive streaming clients,” in Proc. 8th ACM
Multimedia Syst. Conf., Jun. 2017, pp. 76–87.
[4] M. Lu, W. Deng, and Y. Shi. 2018. TF-idletimeout: Improving e ciency of TCAM in SDN by
dynamically adjusting ow entry lifecycle. In Proceedings of the 2016 IEEE International Conference
on Systems, Man, and Cybernetics (SMC). IEEE, Budapest, 2681–2686.
[5] M. H. Mazhar and Z. Shafiq, “Real-time video quality of experience monitoring for HTTPS and
QUIC,” in Proc. IEEE Conf. Comput. Commun. (INFOCOM), Apr. 2018, pp. 1331–1339.
[6] W. Robitza, “Towards behavior-oriented quality of experience assessment for online video services,”
in Proc. Int. Conf. Interact. Experiences TV Online Video, Jun. 2017, pp. 123–127.
[7] Tam N. Nguyen, 2018 2nd Cyber Security in Networking Conference (CSNet), The Challenges in
SDN/ML Based Network Security : A Survey. 4 Apr 2018.
[8] Ting Y, Al-Fuquaha A, Shuaib K, M. Sallabi F, Qadir J, 2018, November. SDN Flow Entry
Management Using Reinforcement Learning, In 2018 Transactions on Autonomous and Adaptive
Systems (ACM).

30
[9] Seufert. M, Wassermann. S and Casas. P, IEEE COMMUNICATIONS LETTERS,. Considering User
Behavior in the Quality of Experience Cycle: Towards Proactive QoE-Aware Traffic Management,
VOL. 23, NO. 7, JULY 2019.
[10] S. C. Lin, I. F. Akyildiz, P. Wang, and M. Luo. 2016. QoS-aware adaptive routing in multi-layer
hierarchical software de ned networks: A reinforcement learning approach. In Proceedings of the
2016 IEEE International Conference on Services Computing (SCC). IEEE, San Francisco, CA, 25–
33.
[11] S. Wang, H. Liu, P. H. Gomes, and B. Krishnamachari. 2017. Deep reinforcement learning for
dynamic multichannel access. In Proceedings of the International Conference on Computing,
Networking and Communications (ICNC). IEEE, Silicon Valley, 257–265.
[12] S. C. Lin, I. F. Akyildiz, P. Wang, and M. Luo. 2016. QoS-aware adaptive routing in multi-layer
hierarchical softwaredefined networks: A reinforcement learning approach. In Proceedings of the
2016 IEEE International Conference on Services Computing (SCC). IEEE, San Francisco, CA, 25–
33.
[13] POX Controller, [Online]. Available: https://github.com/noxrepo/pox. February 2020.
[14] Google DeepMind. [Online]. Available: https://deepmind.com/ . February 2020.
[15] S.Garba, Metric-Based Framework for Testing & Evaluation of Service-Oriented System,
Intenational Journal of Software Engineering & Application (IJSEA) , Vol.10, No.3, May 2019.
[16] L.Zeheng and H.LiGuo, TRACING REQUIREMENTS AS A PROBLEM OF MACHINE
LEARNING, International Journal of Software Engineering & Applications (IJSEA), Vol.9, No.4,
July 2018.

Learning-based Orchestrator for Intelligent Software-defined Networking Controllers

More Related Content

What's hot (17)

Similar to Learning-based Orchestrator for Intelligent Software-defined Networking Controllers (20)

Recently uploaded (20)

Learning-based Orchestrator for Intelligent Software-defined Networking Controllers