Page 141 - Proceedings of the 2017 ITU Kaleidoscope

MACHINE LEARNING APPROACH FOR QUALITY ADAPTATION OF STREAMING
VIDEO THROUGH 4G WIRELESS NETWORK OVER HTTP

1
3
1
2
Dhananjay Kumar , Aswini V ., Arun Raj L ., and Hiran Kumar S .
1 Department of Information Technology, Anna University, MIT Campus, Chennai
2 Department of Computer Science and Engineering, B.S.A Crescent University, Chennai
3 Department of Electronics and Communications Engineering, Vel Tech University, Chennai
dhananjay@annauniv.edu, sindhu16895@gmail.com, arun4u85mit@gmail.com, hksbxr@gmail.com

ABSTRACT The widely used Hypertext Transfer Protocol (HTTP)
employing Transmission Control Protocol (TCP) has
Video streaming over HTTP through 4G wireless network become a primary approach for supporting adaptive video
used for multimedia applications faces many challenges streaming over internet [4].The Dynamic Adaptive
due to fluctuations in network conditions. The existing Streaming over HTTP (DASH) [5] has turned into the
HTTP Adaptive Streaming (HAS) techniques based on accepted video transport mechanism these days, which
prediction of buffer state or link bandwidth offer solution to exploits the functions of widely used HTTP platforms in the
some extent, but if the link condition deteriorates, the internet world.HTTP Adaptive Streaming (HAS) is widely
adaptation process may reduce the streaming bit rate below used model in DASH which encodesa video content with
an acceptable quality level.In this paper,we propose a numerous video qualities (chunks) that have diverse bitrate
machine learning based method, State Action Reward State and adapts the video qualities based on client’s feedback
Action (SARSA) Based Quality Adaptation algorithm using [6].
Softmax Policy (SBQA-SP), which identifies the current In HAS based implementation, the selection of chunk
state (Throughput), action (Streaming quality) and reward duration directly effects the bit rate adaptation process. For
(current video quality) at client to determine the future state example, a small chunk leads to a sub-optimal
and action of the system. The ITU-T G.1070 implementation, while a larger chunk will cause lack of
recommendation (parametric) modelis embedded in the adaptation in the fast changing internet traffic.The adoption
SBQA-SP to implement adaptation process. The proposed of TCP/HTTP leads to an inefficient network bandwidth
system was implemented on the top of HTTP in a typical utilization, and a mismatch between the specified quality of
internet environment using 4G wireless network and the a chunk and the actual encoding rate further aggravate the
streaming quality is analyzed using several full reference problem [7].
video metrics. The test results outperformed the existing Q- The bit rate adaptation algorithm needs to deal with multi-
Learning based video quality adaptation (QBQA) dimensional aspects of video streaming over HTTP through
algorithm. For instance, an improvement of 5% in average wireless networks. Most video codec, e.g., High Efficiency
PSNR and 2 % increase in average SSIM index over the Video Coding (HEVC), H.264/AVC etc., generates variable
QBQA approach was observed for the live stream. bit rate of encoded video. However, the meta-data of
MPEG-DASH does not carry this which can be used by the
Keywords— HTTP Adaptive Streaming, SARSA, Q- client for adaptation process [8]. The existing HAS
Learning, Video Quality Adaptation. approach do not provide control of transfer rate of video
data. In fact, the TCP controls the transmission rate of
1. INTRODUCTION video chunk, which respond to the congestion in network
connecting client and server [9]. The fluctuation in received
The explosive growth of multimedia application and signal strength in wireless network further inflicts the
services over wired and wireless networks demands a system capacity. In a typical multiple access cellular
paradigm shift in system design. In North America, 71% system, the data rate at user equipment depends on
percent of data traffic in evening consists of streaming prevailing channel conditions[10].Most of the earlier work
audio and video over fixed networks [1]. Further, as per the tries to estimate the future bandwidth and hence the
Cisco Visual Networking Index (VNI), the video in IP efficiency of this approach depends on accuracy of
network will account for 82% of total consumer internet prediction. However, it is inherentlydifficult to predict the
traffic globally in 2021 [2]. It is also reported that in the receiving bit rate based on past history [11].
year 2016, the Fourth Generation (4G) mobile system A machine learning technique can be employed in
supported 69% of the total mobile traffic [3]. This adaptation process provided it is incorporated into feedback
extremely fast increase in video data traffic poses a quality loop. Reinforcement Learning (RL)[12]is an
challenge to network operators and system designers to efficient solution for environmental learning problem. In
enhance the abilityof the networksconsidering end-to-end RL, rather than relying on a fixed algorithm, learning
connectivitywhilesupporting high-quality video streaming.
agents can try different actions and gradually learn the best

978-92-61-24291-6/CFP1768P-ART © 2017 ITU – 125 – Kaleidoscope