Machine learning in Dynamic Adaptive Streaming over HTTP (DASH)

Int. J. Advanced Networking and Applications
Volume: 09 Issue: 03 Pages: 3461-3468 (2017) ISSN: 0975-0290
3461
Machine learning in Dynamic Adaptive Streaming
over HTTP (DASH)
Koffka Khan
Department of Computing and Information Technology The University of the West Indies, Trinidad and Tobago, W.I
Email: koffka.khan@gmail.com
Wayne Goodridge
Department of Computing and Information Technology The University of the West Indies, Trinidad and Tobago, W.I
Email: wayne.goodridge@sta.uwi.edu.com
-------------------------------------------------------------------ABSTRACT---------------------------------------------------------------
Recently machine learning has been introduced into the area of adaptive video streaming. This paper explores a
novel taxonomy that includes six state of the art techniques of machine learning that have been applied to
Dynamic Adaptive Streaming over HTTP (DASH): (1) Q-learning, (2) Reinforcement learning, (3) Regression, (4)
Classification, (5) Decision Tree learning, and (6) Neural networks.
Keywords—Machine learning; DASH; Q-learning; Reinforcement learning; Regression; Classification; Decision
Tree learning; Neural networks
---------------------------------------------------------------------------------------------------------------------------------------------------
Date of Submission: Nov 17, 2017 Date of Acceptance: Dec 05, 2017
---------------------------------------------------------------------------------------------------------------------------------------------------
I. INTRODUCTION
The growing trend of the usage of video services
demands the availability of efficient and reliable video
quality assessment and transmission techniques. In a HAS
design, video content is stored on a server as segments of
fixed duration at different quality levels. Each client can
request the segment at the most appropriate quality level on
the basis of the local perceived bandwidth or its internal
buffer level. In this way, video playback dynamically
changes according to the available resources, resulting in a
smoother video streaming.
Typical goals that measure the goodness of a streaming
session take into account average quality levels displayed
to the user, fluctuations in quality levels, and rebuffering
events [21]. Unfortunately, there is no single agreed upon
measurement of streaming quality. The ultimate standard of
goodness might involve conducting comprehensive user
surveys under different network conditions and devices
[37]. Hence, the ability to optimize with respect to a wide
variety of objective functions that take these events into
account is of utmost importance.
Q-learning can also be viewed as a method of
asynchronous dynamic programming (DP). It provides
agents with the capability of learning to act optimally in
Markovian domains by experiencing the consequences of
actions, without requiring them to build maps of the
domains [35]. In Q-learning, if each subagent uses the
conventional Q-learning algorithm [39], global optimality
is not achieved. Instead, each subagent learns the Qj values
that would result if that subagent were to make all future
decisions for the agent. This “illusion of control” means
that the subagents converge to “selfish” estimates that
overestimate the true values of their own rewards with
respect to a globally optimal policy [24].
In adaptive streaming environments the concept of
reinforcement learning (RL) is at client side. The client is
able learn the most appropriate parameter configuration
under different network conditions. RL is a machine
learning technique, in which an agent can learn about its
environment by performing a number of actions. Every
time an action is taken, the agent perceives feedback
through a numerical reward from the environment [30].
The agent’s goal is to learn which action should be taken in
a given environmental state, in order to maximize the
cumulative numerical reward.
Regression is a simple, yet powerful technique for
identifying trends in datasets and predicting properties of
new data points. The output can be a continuous value, a
discrete value (e.g. a boolean), or a vector of such values.
The input is a vector, and can also be either discrete or
continuous. The individual properties of one data row are
called features. For discrete problems, threshold value(s)
is/are chosen between the various “levels” of output. The
phase-space surface defined by these levels is called the
decision boundary [14]. The prediction is done via a
hypothesis function. The “learning” is driven by a set of
examples.
Classification is the procedure of placing a new
observation into a set of categories (sub-categories), on the
basis of a training set of data, which contains observations
(or instances) whose category membership is known [18].
The classification model can be developed on the complete
case data for a given feature. This model treats the feature
as the outcome and uses the other features as predictors.
Feature subset selection is the process of identifying and
removing as many irrelevant and redundant features as
possible [12]. This reduces the dimensionality of the data
and enables data mining algorithms to operate faster and
more effectively.
The decision tree is defined as a supervised learning
model that hierarchically maps a data domain onto a
response set. It divides a data domain (node) recursively
into two sub domains such that the subdomains have a
higher information gain than the node that was split. The
goal of supervised learning is the classification of the data,
and therefore, the information gain means the ease of
classification in the subdomains created by a split. Finding

3462
the best split that gives the maximum information gain (i.e.,
the ease of classification) is the goal of the optimization
algorithm in the decision tree-based supervised learning
[29].
Artificial neural networks can be employed to solve a
wide spectrum of problems in optimization, parallel
computing, matrix algebra and signal processing [6]. There
exist many training methods for both feedforward networks
(including multilayer and radial basis networks) and
recurrent networks. In addition to conjugate gradient and
Levenberg-Marquardt variations of the backpropagation
algorithm (BA), the Neural Network BAs also covers
Bayesian regularization and early stopping, which ensure
the generalization ability of trained networks [8]. Practical
uses of neural networks includes training for function
approximation, pattern recognition, clustering, prediction,
and of lately adaptive video streaming.
This work consists of four sections. Section II presents
the Taxonomy of Machine Learning-based techniques
currently applied to DASH. Section III gives examples of
these state of the art techniques that were defined in the
taxonomy. Finally, the conclusion is given in Section IV.
II. TAXONOMY OF MACHINE LEARNING-BASED
AVS
The Machine Learning-based AVS techniques are
categorized into (1) Q-Learning, (2) Reinforcement
Learning, (3) Regression, (4) Classification, (5) Decision
Tree Learning, and (6) Neural Networks. (cf. Figure 1)
Figure 1: Machine Learning in AVS: a taxonomy.
III. MACHINE-LEARNING IN DASH
A. Q-Learning
An adaptive Q-Learning-based HAS client is proposed in
[4]. The proposed HAS client dynamically learns the
optimal behavior corresponding to the current network
environment. This is in contrast to the way current
heuristics deal with the AVS problem. It considers (1)
multiple aspects of video quality, (2) a tunable reward
function which gives the opportunity to focus on different
aspects of the Quality of Experience (QoE), and (3) the
quality as perceived by the end-user. The proposed HAS
client is evaluated by a network-based simulator. It
investigates multiple reward configurations and Q-learning
specific settings. The proposed client outperforms standard
HAS players in the evaluated networking environments.
Figure 2: Schematic overview of the Q-learning HAS client
design. [4]
Authors in [31] use a novel learning model to determine
(a) the optimal time to re-route the traffic flows, and (b) to
change the bitrate of the video. To accomplish this task
they propose an adaptive video streaming system with a
learning-based approach. It runs over a Software Defined
Network (SDN) [11], [13]. In the proposed video
streaming system the learning model aims to minimize (1)
the packet loss rate, (2) quality changes, and (3) controller
cost. It concurrently adapts (i) the flow routes and (ii)
video quality. The performance of the learning-based
approach is tested by comparing it to traditional Internet
routing and the greedy approach. The proposed system
significantly outperforms the traditional Internet routing
approach and the greedy approach in experiments under
different network scenarios. The tests covered both
Quality of Experience (QoE) and network cost.
Figure 3: Network topology used in the simulations. There
are six clients, one server and one controller in the
network. The blue dotted lines indicate the OpenFlow
protocol and the red dashed line represents the
communication between the server and the controller. [31]
A (frequency adjusted) Q-learning HAS client is proposed
in [5]. The proposed HAS client dynamically learns the

3463
optimal behaviour in terms of the Quality of Experience
(QoE) corresponding to the current network environment.
Thus, the client has been optimised both in terms of (1)
global performance and (2) convergence speed. In a wide
range of network configurations, thorough evaluations
show that the proposed client can outperform deterministic
algorithms by 11–18% in terms of mean opinion score
[28], [36] (MOS).
Figure 4: Overview of the simulated topology. [5]
Authors in [17] propose a Q-Learning-based algorithm for
an HTTP Adaptive Streaming (HAS). It maximizes the
perceived user quality. Account is taken into the
relationships amongst (a) the estimated bandwidth, (2) the
qualities, and (3) penalizing the freezes. Results show that
it produces an optimal control as observed with other
approaches do, but in addition it keeps the adaptiveness.
B. Reinforcement Learning
Authors in [22] propose a Reinforcement Learning-based
quality selection algorithm. It is able to achieve fairness in
a multi-client setting. A coordination proxy in charge of
facilitating the coordination among clients is a key
element of their approach. There are three advantages to
this approach: (1) the algorithm is able to learn and adapt
its policy depending on network conditions. This is unlike
recently developed HAS heuristics. (2) fairness is
achieved without explicit communication among agents.
Thus there is no significant overhead introduced into the
network. (3) There are no modifications to the standard
HAS architecture. This novel approach is evaluated
through simulations using (a) mutable network conditions,
and (b) in several multi-client scenarios. The proposed
approach improves fairness by up to 60% compared to
current HAS heuristics.
Figure 5: Logical work-flow of the proposed solution. [22]
An adaptive method that combines video encoding and the
video transmission control system is presented in [2]. The
study was implemented over heterogeneous networks. The
method includes the following steps: (1) collect and
standardize the real-time information describing the
network and the video, (2) assess the video quality, (3)
calculate the video coding rate based on the standardized
information, (4) process the encoded compression of the
video according to the calculated coding rate, and (5)
transfer the compressed video. It was shown by
experiments that there is a significant improvement for the
quality of real-time videos transmission. This is achieved
without changing the existing network. The advantages of
the solution are (1) it is easy to deploy. (2) easy to
implement quickly, and (3) may help to extensively ensure
video quality for normal users.
Figure 6: Proposed system framework. [2]
A reinforcement learning approach that uses only the
buffer state information to choose the segment quality
during playback is proposed in [44]. This approach
optimizes for a measure of user-perceived streaming
quality. Results are simulated and show that the proposed
approach achieves better QoE than rate-based, buffer-
based, and reinforcement learning-based approaches.
The concept of reinforcement learning is introduced at
client side in [32]. It allows the player to adaptively
change the parameter configuration for existing rate
adaptation heuristics. Characteristics that are taken into
account in the decision process are bandwidth-based. This
allows the player to improve its bandwidth-awareness. The
goals of the model are (1) to actively reduce the average
buffer filling, and (2) to evaluate the results for two well-
known heuristics: (a) the Microsoft IIS Smooth Streaming
heuristic [43], and (2) the QoE-driven Rate Adaptation
Heuristic for Adaptive video Streaming [22]. The average
buffer filling can be reduced by 8.3% compared to state of
the art by using the proposed learning-based approach.
This reduction is achieved while achieving a comparable
level of QoE.
Authors in [16] develop Pensieve. Pensieve is a system
that generates ABR algorithms. It only uses
Reinforcement Learning (RL) techniques. Pensieve uses
RL to train a neural network model. The model selects
bitrates for future video chunks. This selection is based on
observations collected by client video players. Pensieve
does not rely upon pre-programmed models or
assumptions about the environment, which makes it
different from other approaches. Instead, it learns to make
ABR decisions solely through observations. These
observations are as a result of the performance of past
decisions made by the controller. Consequently, Pensieve
can automatically allow ABR algorithms that adapt to (1)

3464
a wide range of environmental conditions, and (2) QoE
metrics. Pensieve is compared to state-of-the-art ABR
algorithms. Pensive uses (a) trace-driven experiments, (b)
real world experiments ((a) and (b) span a wide variety of
network conditions), (c) QoE metrics, and (d) video
properties. Pensieve outperforms the best state of the art
scheme in all considered scenarios. It obtains
improvements in average QoE of 13.1%-25.0%. Pensieve's
policies generalize well. It outperforms existing schemes.
This is so even on networks on which it was not trained.
Figure 7: The Actor-Critic algorithm that Pensieve uses to
generate
ABR policies. [16]
Authors in [41] discuss the use of reinforcement-learning
(RL) to learn the optimal request strategy at the HAS
client. It progressively maximizing a pre-defined Quality
of Experience (QoE)-related reward function. The most
influential factors for the request strategy are investigated
under a RL framework. The strategy uses a forward
variable selection algorithm. The performance of the RL-
based HAS client is evaluated. It is done on a Video-on-
Demand (VOD) simulation system. Results are gathered
from experiments. The RL-based HAS client is able to
optimize the quantitative QoE given the QoE-related
reward function. The RL-based HAS client is more robust
and flexible compared to a conventional HAS system,
even under versatile network conditions.
C. Regression
In is shown in [15] that static users and adaptive streaming
users have less starvation events than mobile users. Also
researchers observed that with mobile users it is more
difficult to predict their video starvation. It was shown that
(1) channel conditions, and (2) number of active users are
two important features which contribute to better
prediction performance. The two features prediction
strategy provides sufficient accuracy for static users.
However, but not sufficient for mobile users. We also
demonstrate that the two information, number of users
served in a cell and the number of users experiencing
video starvation, provide similar prediction accuracy.
In the proposed method [19], authors use the QoS
parameters for the video streaming. These parameters are
determined based on a machine learning algorithm. It uses
regression analysis. The parameters used for the model are
(a) according to the user requirements, (b)
computational/network resources, and (c) service
provisioning environments. The design and
implementation of a prototype shows the method is
experimentally valid and feasible.
Figure 8: The Video quality determining mechanism with
machine learning. [19]
Authors in [27] investigate quality prediction models for
videos. They use (1) well-known image metrics, (2)
information about the bitrate levels, and (3) a relatively
simple machine learning method. Even though quality
assessment of ABR videos is a hard problem, the research
shows that initial results are promising. A Spearman rank
order correlation of 0.88 is obtained using content-
independent cross-validation.
A no-reference objective quality estimation framework is
proposed in [26]. The framework is suitable for any block-
based video codec. In the proposed solution, features are
extracted from coding units. They are then summarized to
form features at frame levels. Stepwise regression is used
to select the important feature variables. This reduces the
dimensionality of feature vectors. Subsequently, a
polynomial regression-based approach is used to model
the nonlinear relationship amid (a) the feature vectors and
(b) the true objective quality values. Such values are likely
for coding units and video frames. The proposed
framework is implemented using MPEG-2 and HEVC.
The objective quality estimation results are compared
against an existing state-of-the-art solution. It is then
quantified using the Pearson correlation factor and the root
mean square error measure.
D. Classification
Authors in [23] propose a network-based framework. A
network controller prioritizes the delivery of particular
video segments. This is done to prevent freezes at the
clients. This framework is based on OpenFlow. OpenFlow
is a widely adopted protocol to implement the software-
defined networking (SDN) principle. The main element of
the controller is a Machine Learning (ML) engine. This
engine is based on the random undersampling boosting
algorithm and fuzzy logic. It can detect when a client is
close to a freeze and drive the network prioritization to
avoid it. This decision is based on measurements collected
from the network nodes only. The engine does not have
any knowledge on (1) the streamed videos or (2) on the

3465
clients' characteristics. The design of the proposed ML-
based framework is discussed and illustrated. In addition,
its performance with other benchmarking HAS solutions is
compared. This is undertaken with various video
streaming scenarios. Principally, through extensive
experimentation it is shown that the proposed approach
can reduce video freezes and freeze time with about 65%
and 45% respectively.
A Machine Learning-based Adaptive Streaming over
HTTP (MLASH) is presented in [3]. It is an elastic
framework and exploits a wide range of useful network-
related features to train a rate classification model.
MLASH allows its machine learning-based framework to
be incorporated with any existing adaptation algorithm and
utilize big data characteristics. This would improve a
model’s prediction accuracy. Trace-based simulations
show that machine learning-based adaptation achieves a
better performance than traditional adaptation algorithms.
Performance is measured in terms of a players target
quality of experience (QoE) metrics.
Figure 9: System Architecture ofMLASH. [3]
Authors in [20] present a methodology for the
classification of end users’ QoE. This is for users watching
YouTube videos, and is based only on statistical properties
of encrypted network traffic. The developed system is
called YouQ. It includes (1) tools for monitoring, (2)
analysis of application-level quality indicators, and (3)
corresponding traffic traces. Collected data is then used for
the development of machine learning models for QoE
classification. This is based on computed traffic features
per video session. A collected dataset corresponding to
1060 different YouTube videos streamed across 39
different bandwidth scenarios is used to test the YouQ
system and methodology. Various classification models
are also tested. Classification accuracy was found to be up
to 84%. Three QoE classes are used (a) “low”, (b)
“medium” or (c) “high”. Up to 91% classification
accuracy is obtained when using binary classification
classes (i) “low” and (2) “high”. Why and when prediction
errors occur are discussed in an attempt to improve the
models in the future.
E. Decision Tree Learning
Authors in [1] tackles the problem of QoE monitoring,
assessment and prediction in cellular networks, relying on
end-user device (i.e., smart-phone) QoS passive traffic
measurements and QoE crowdsourced feedback. We
conceive different QoE assessment models based on
supervised machine learning techniques, which are
capable to predict the QoE experienced by the end user of
popular smartphone apps (e.g., YouTube and Facebook),
using as input the passive in-device measurements. Using
a rich QoE dataset derived from field trials in operational
cellular networks, we benchmark the performance of
multiple machine learning based predictors, and construct
a decision-tree based model which is capable to predict the
per-user overall experience and service acceptability with
a success rate of 91% and 98% respectively To the best of
our knowledge, this is the first paper using end-user, in-
device passive measurements and machine learning
models to predict the QoE of smartphone users in
operational cellular networks.
Authors in [7] have used the Institut National de la
Recherche Scientifique (INRS) audiovisual quality dataset
which is specifically designed to include ranges of
parameters and degradations typically seen in real-time
communications to build Decision Trees--based Ensemble
(DTE) methods. It is shown that DTEs have outperformed
both Deep Learning-based and Genetic Programming-
based models. This is in terms of the (a) Root-Mean-
Square Error (RMSE), and (b) Pearson correlation values.
It was shown through other conducted experiment that
Random Forests-based prediction models achieve high
accuracy for both the INRS audiovisual quality dataset and
other publicly available comparable datasets.
F. Neural Networks
A novel automated and computationally efficient video
assessment method is introduced in [34]. The method
enables accurate real-time (online) analysis of delivered
quality in an adaptable and scalable manner. Offline deep
unsupervised learning processes are employed at the
server side. The method is inexpensive for the client as
there is a no-reference measurements at the client side. It
provides both real-time assessment and performance
comparable to the full reference counterpart at the client
side. The method is tested on the LIMP Video Quality
Database [33]. It obtains a correlation between 78% and
91% to the FR benchmark [38]. The method is flexible,
dynamically adaptable to new content and is scalable with
the number of videos because of its unsupervised learning
essence.
Figure 10: Real-time UDL-based video quality assessment
method. [34]
Authors in [10] present D-DASH. This is a framework that
combines Deep Learning and Reinforcement Learning
techniques. It optimizes the Quality of Experience (QoE)
of DASH. Authors propose different learning
architectures. They are assessed, combined with feed-

3466
forward and recurrent deep neural networks and
incorporates advanced strategies. D-DASH designs are
thoroughly evaluated against state-of-the-art, heuristic and
learning-based algorithms. Performance indicators such as
image quality across video segments and
freezing/rebuffering events are used.
Results are obtained on both real and simulated channel
traces. They show the superiority of D-DASH in nearly all
the considered quality metrics. The D-DASH framework
exhibits faster convergence to the rate-selection strategy
than the other learning algorithms considered in the study
and yields a considerably higher QoE.
Figure 11: Network architectures. [10]
Authors in [25] present bitstream-based features for
perceptual quality estimation of HEVC coded videos.
Various factors are used including: (1) the impact of
different sizes of block-partitions, (2) use of reference-
frames, (3) the relative amount of various prediction
modes, (4) statistics of motion vectors and, (5)
quantization parameters are taken into consideration. The
result are 52 features relevant for perceptual quality
prediction being produced. The used test stimuli
constitutes 560 bitstreams. These have been carefully
extracted for this analysis from the 59, 520 bistreams of
the large-scale database generated by the Joint Effort
Group (JEG) [9]. The significance of the considered
features through reasonably accurate and monotonic
prediction of a number of objective quality metrics is
highlighted.
Authors in [42] propose a novel no-reference (NR) video
quality metric that evaluates the impact of frame freezing
due to either packet loss or late arrival. The metric uses a
trained neural network (NN). The NN acts on features that
are chosen to capture the impact of frame freezing on the
user perceived quality. The considered features include (a)
the number of freezes, (b) freeze duration statistics, (c)
inter-freeze distance statistics, (d) frame difference before
and after the freeze, (e) normal frame difference, and (f)
the ratio of them. The NN finds the mapping between
features and subjective test scores. It optimizes the
network structure and the feature selection. This is done
through a cross-validation procedure. The cross-validation
uses training samples extracted from both VQEG [40] and
LIVE video databases. The resulting feature set and
network structure gives precise quality prediction for both
the training data containing 54 test videos and an
unconnected testing dataset including 14 videos, with
Pearson correlation coefficients greater than 0.9 and 0.8
for the training set and the testing set, respectively.
Figure 12: Flowchart of the proposed metric. [42]
IV. CONCLUSION
Recently machine learning has been introduced into the
area of adaptive video streaming. This paper explores a
novel taxonomy that includes six state of the art techniques
of machine learning that have been applied to Dynamic
Adaptive Streaming over HTTP (DASH): (1) Q-learning,
(2) Reinforcement learning, (3) Regression, (4)
Classification, (5) Decision Tree learning, and (6) Neural
networks.
REFERENCES
[1] Casas, Pedro, Alessandro D'Alconzo, Florian
Wamser, Michael Seufert, Bruno Gardlo, Anika
Schwind, Phuoc Tran-Gia, and Raimund Schatz.
"Predicting QoE in cellular networks using machine
learning and in-smartphone measurements." In
Quality of Multimedia Experience (QoMEX), 2017
Ninth International Conference on, pp. 1-6. IEEE,
2017.
[2] Cheng, Bo, Jialin Yang, Shangguang Wang, and
Junliang Chen. "Adaptive video transmission control
system based on reinforcement learning approach
over heterogeneous networks." IEEE Transactions on
Automation Science and Engineering 12, no. 3
(2015): 1104-1113.
[3] Chien, Yu-Lin, Kate Ching-Ju Lin, and Ming-Syan
Chen. "Machine learning based rate adaptation with
elastic feature selection for HTTP-based streaming."
In Multimedia and Expo (ICME), 2015 IEEE
International Conference on, pp. 1-6. IEEE, 2015.
[4] Claeys, Maxim, Steven Latré, Jeroen Famaey,
Tingyao Wu, Werner Van Leekwijck, and Filip De
Turck. "Design of a Q-learning-based client quality
selection algorithm for HTTP adaptive video
streaming." In Adaptive and Learning Agents
Workshop, part of AAMAS2013 (ALA-2013), pp. 30-
37. 2013.
[5] Claeys, Maxim, Steven Latré, Jeroen Famaey,
Tingyao Wu, Werner Van Leekwijck, and Filip De
Turck. "Design and optimisation of a (FA) Q-
learning-based HTTP adaptive streaming client."
Connection Science 26, no. 1 (2014): 25-43.
[6] Cochocki, A., and Rolf Unbehauen. Neural networks
for optimization and signal processing. John Wiley &
Sons, Inc., 1993.
[7] Demirbilek, Edip, and Jean-Charles Grégoire.
"Machine Learning--Based Parametric Audiovisual
Quality Prediction Models for Real-Time
Communications." ACM Transactions on Multimedia
Computing, Communications, and Applications
(TOMM) 13, no. 2 (2017): 16.

3467
[8] Demuth, Howard B., Mark H. Beale, Orlando De Jess,
and Martin T. Hagan. Neural network design. Martin
Hagan, 2014.
[9] Fliegel, Karel. "QUALINET Multimedia Databases
v4. 6." (2013).
[10]Gadaleta, Matteo, Federico Chiariotti, Michele Rossi,
and Andrea Zanella. "D-DASH: a Deep Q-learning
Framework for DASH Video Streaming." IEEE
Transactions on Cognitive Communications and
Networking (2017).
[11]Kirkpatrick, Keith. "Software-defined networking."
Communications of the ACM 56, no. 9 (2013): 16-19.
[12]Kotsiantis, Sotiris B., Ioannis D. Zaharakis, and
Panayiotis E. Pintelas. "Machine learning: a review of
classification and combining techniques." Artificial
Intelligence Review 26, no. 3 (2006): 159-190.
[13]Kreutz, Diego, Fernando MV Ramos, Paulo Esteves
Verissimo, Christian Esteve Rothenberg, Siamak
Azodolmolky, and Steve Uhlig. "Software-defined
networking: A comprehensive survey." Proceedings
of the IEEE 103, no. 1 (2015): 14-76.
[14]Kutner, Michael H., Chris Nachtsheim, and John
Neter. Applied linear regression models. McGraw-
Hill/Irwin, 2004.
[15]Lin, Yu-Ting, Eduardo Mucelli Rezende Oliveira,
Sana Ben Jemaa, and Salah Eddine Elayoubi.
"Machine learning for predicting QoE of video
streaming in mobile networks." In Communications
(ICC), 2017 IEEE International Conference on, pp. 1-
6. IEEE, 2017.
[16]Mao, Hongzi. "Neural adaptive video streaming with
pensieve." PhD diss., Massachusetts Institute of
Technology, 2017.
[17]Martín, Virginia, Julián Cabrera, and Narciso García.
"Evaluation of q-learning approach for HTTP
adaptive streaming." In Consumer Electronics
(ICCE), 2016 IEEE International Conference on, pp.
293-294. IEEE, 2016.
[18]Michie, Donald, David J. Spiegelhalter, and Charles
C. Taylor. "Machine learning, neural and statistical
classification." (1994).
[19]Oide, Makoto, Akiko Takahashi, Toru Abe, and
Takuo Suganuma. "Design and implementation of
user-oriented video streaming service based on
machine learning." In Cognitive Informatics &
Cognitive Computing (ICCI* CC), 2016 IEEE 15th
International Conference on, pp. 111-116. IEEE,
2016.
[20]Orsolic, Irena, Dario Pevec, Mirko Suznjevic, and Lea
Skorin-Kapov. "A machine learning approach to
classifying YouTube QoE based on encrypted
network traffic." Multimedia Tools and Applications
(2017): 1-35.
[21]Oyman, Ozgur, and Sarabjot Singh. "Quality of
experience for HTTP adaptive streaming services."
IEEE Communications Magazine 50, no. 4 (2012).
[22]Petrangeli, Stefano, Maxim Claeys, Steven Latré,
Jeroen Famaey, and Filip De Turck. "A multi-agent
Q-Learning-based framework for achieving fairness
in HTTP Adaptive Streaming." In Network
Operations and Management Symposium (NOMS),
2014 IEEE, pp. 1-9. IEEE, 2014.
[23]Petrangeli, Stefano, Tingyao Wu, Tim Wauters,
Rafael Huysegems, Tom Bostoen, and Filip De Turck.
"A machine learning-based framework for preventing
video freezes in HTTP adaptive streaming." Journal of
Network and Computer Applications 94 (2017): 78-
92.
[24]Russell, Stuart J., and Andrew Zimdars. "Q-
decomposition for reinforcement learning agents." In
Proceedings of the 20th International Conference on
Machine Learning (ICML-03), pp. 656-663. 2003.
[25]Shahid, Muhammad, Joanna Panasiuk, Glenn Van
Wallendael, Marcus Barkowsky, and Benny
Lövström. "Predicting full-reference video quality
measures using HEVC bitstream-based no-reference
features." In Quality of Multimedia Experience
(QoMEX), 2015 Seventh International Workshop on,
pp. 1-2. IEEE, 2015.
[26]Shanableh, Tamer. "A regression-based framework
for estimating the objective quality of HEVC coding
units and video frames." Signal Processing: Image
Communication 34 (2015): 22-31.
[27]Søgaard, Jacob, Søren Forchhammer, and Kjell
Brunnström. "Quality assessment of adaptive bitrate
videos using image metrics and machine learning." In
Quality of Multimedia Experience (QoMEX), 2015
Seventh International Workshop on, pp. 1-2. IEEE,
2015.
[28]Streijl, Robert C., Stefan Winkler, and David S.
Hands. "Mean opinion score (MOS) revisited:
methods and applications, limitations and
alternatives." Multimedia Systems 22, no. 2 (2016):
213-227.
[29]Suthaharan, Shan. "Decision tree learning." In
Machine Learning Models and Algorithms for Big
Data Classification, pp. 237-269. Springer US, 2016.
[30]Sutton, Richard S., and Andrew G. Barto.
Reinforcement learning: An introduction. Vol. 1, no.
1. Cambridge: MIT press, 1998.
[31]Uzakgider, Tuba, Cihat Cetinkaya, and Muge Sayit.
"Learning-based approach for layered adaptive video
streaming over SDN." Computer Networks 92 (2015):
357-368.
[32]van der Hooft, Jeroen, Stefano Petrangeli, Maxim
Claeys, Jeroen Famaey, and Filip De Turck. "A
learning-based algorithm for improved bandwidth-
awareness of adaptive streaming clients." In
Integrated Network Management (IM), 2015
IFIP/IEEE International Symposium on, pp. 131-138.
IEEE, 2015.
[33]Vega, M. Torres, and A. Liotta. "LIMP-video quality
database." (2016).
[34]Vega, Maria Torres, Decebal Constantin Mocanu,
Jeroen Famaey, Stavros Stavrou, and Antonio Liotta.
"Deep Learning for Quality Assessment in Live Video
Streaming." IEEE Signal Processing Letters 24, no. 6
(2017): 736-740.

3468
[35]Vidal, José M. "Learning in multiagent systems: An
introduction from a game-theoretic perspective." In
Adaptive agents and multi-agent systems, pp. 202-
215. Springer, Berlin, Heidelberg, 2003.
[36]Viswanathan, Mahesh, and Madhubalan Viswanathan.
"Measuring speech quality for text-to-speech systems:
development and assessment of a modified mean
opinion score (MOS) scale." Computer Speech &
Language 19, no. 1 (2005): 55-83.
[37]Wang, Zhiheng, Sujata Banerjee, and Sugih Jamin.
"Studying streaming video quality: from an
application point of view." In Proceedings of the
eleventh ACM international conference on
Multimedia, pp. 327-330. ACM, 2003.
[38]Wang, Zhou, Ligang Lu, and Alan C. Bovik. "Video
quality assessment based on structural distortion
measurement.” Signal processing: Image
communication 19, no.2(2004):121-132.
[39]Watkins, Christopher JCH, and Peter Dayan. "Q-
learning." Machine learning 8, no. 3-4 (1992): 279-
292.
[40]Winkler, Stefan. “Analysis of public image and video
database for quality assessment.” IEEE Journal of
selected topics in Signal Processing 6, no. 6 (2012):
616-625
[41]Wu, Tingyao, and Wemer Van Leekwijck. “Factor
selection for reinforcement learning in http adaptive
streaming.” In International Conference on
Multimedia Modeling, pp.553-567. Springer, Cham,
2014.
[42]Xue, Yuanyi, Beril Erkin, and Yao Wang. "A novel
no-reference video quality metric for evaluating
temporal jerkiness due to frame freezing." IEEE
Transactions on Multimedia 17, no.1 (2015):134-139.
[43]Zambelli, Alex. "IIS smooth streaming technical
overview." Microsoft Corporation 3 (2009):40.
[44]Zhang, Yue, and Yao Liu. "Buffer-Based
Reinforcement Learning for Adaptive Streaming." In
Distributed Computing Systems(ICDCS), 2017 IEEE
37th
International Conference on, pp.2569-2570. IEEE,
2017.
Biographies and Photographs
Koffka Khan received the M.Sc., and
M.Phil. degree from the University of the
West Indies. He is currently a PhD student
and has up-to-date, published numerous
papers in Journals and proceedings of
international repute. His research areas are computational
intelligence, routing protocols, wireless communications,
information security and adaptive streaming controllers.
Wayne Goodridge is a Lecturer in the
Department of Computing and Information
Technology, The University of the West
Indies, St. Augustine. He did is PhD at
Dalhousie University and his research interest includes
computer communications and security.

Machine learning in Dynamic Adaptive Streaming over HTTP (DASH)

More Related Content

What's hot (19)

Similar to Machine learning in Dynamic Adaptive Streaming over HTTP (DASH) (20)

More from Eswar Publications (20)

Recently uploaded (20)

Machine learning in Dynamic Adaptive Streaming over HTTP (DASH)