Page 142 - Proceedings of the 2017 ITU Kaleidoscope

2017 ITU Kaleidoscope Academic Conference

strategy for each situation. By continuous learning the RL connectivity between client and server is supported on a 4G
algorithms like Q-learning can adapt to the changing wireless network. Initially, the client requests the Media
conditions of the streaming system.However, the Descriptor file from the server, and the server replies with
complexity of the model based on Q-learning [13] could the Media Descriptor Sidecar file containing default
seriously downgrade the system performance especially in settings and video parameters. Once the server streams the
dealing with the live streaming of video. video to the client, the client continuously monitors the
The Full Reference (FR) metrics [14] of video quality streaming quality using proposed SARSA based quality
evaluation produces best result as it compares the received adaptation algorithm to determine the corrective action to
signal with original at frame level. However using FR be taken by the server in the near future and send this
metrics in dynamic adaptation of quality is not practical as decision as feedback to the server. The server adapts the
the receiver does not have the original video. If the learning streaming video quality accordingly to match the client’s
technique can be incorporated into No Reference requirement.
(NR)metrics of video quality estimation, a dynamic
streaming system can be designed and developed to deal Table 1.Test Parameters as per ITU-T J.247
with the client’s terminal requirement.Although, the ITU-T
G.1070 [15] recommendation is targeted towards quality of S.no Parameters Values
experience / service (QoE/QoS) planners in video 1. Transmission Errors with packet loss
telephony, its parametric model is adapted here in 2. Frame Rate 5 fps to 30 fps
supporting video streaming system to meet the end to end 3. Video Codec H.264/AVC (MPEG-4
service quality. part10),VC-1,Windows
In this work, we propose a new algorithm based on RL Media9, Real Video (RV
approach called, State Action Reward State Action 10),MPEG-4 Part 2
(SARSA) Based Quality Adaptation using Softmax Policy 4. Video QCIF: 16 - 320 kbps
(SBQA-SP) algorithm to manage the adaptive streaming Resolutions and CIF: 64 - 2000 kbps
using NR metrics. SARSA is an online policy approach of bit rates VGA: 128 - 4000 kbps
RL [16], which doesn’t require a separate learning and 5. Temporal errors Maximum of 2 seconds
deployment phase. In SBQA-SP,the system is characterised (pausing with
by a set of states and the algorithm decides the suitable skipping)
action to be taken based on the current state. The SBQA-SP
identifies the current state of the system and based on the 2.2. Server Side Functions
state chooses an action to perform. It calculates the reward
as a result of the action performed and determines the The server initially set media locator URL either with
resulting state of the system after the action. Next, the location of the stored video or with the application program
SBQA-SP determines the future action to be performed interface (DirectShow [18]) for accessing camera to capture
based on theSoftmax exploration policy and update the Q- live video. The media player / encoder objectis initialized to
matrix. The chosen action is sent as feedback to the server. performvideo transcoding during streaming process. The
To analyse the performance of the proposed approach, it is variation of the frame rate is set between lower (e.g., 20
compared with other two approaches, namely (i) Q- fps) and upper limit (e.g., 30 fps) and theresolution is set
Learning Based Quality Adaptation (QBQA) [13], and (ii) with standards like QCIF, CIF etc. The destination
SBQA using ε-Greedy Policy (SBQA-GP). In QBQA, Q- addressislinked with server’s IP address and application’s
learning method is used for controlling video quality port number. Once the transcoding parameters are set,the
adaptation. The Q-Learning approachis similar to SARSA, server starts streaming at the specified URL through the
expect for the fact that it is an off-line policy algorithm HTTP port.Itwaits for client’s reply in the application’s port
which requires a learning and deployment phase. Also, the and connect with the client which request for connection.
formula to update Q-matrix varies for SARSA and Q- The server continuously listens for client feedback about
learning. SBQA-GP is a variant of SBQA-SP approach in the video quality, andthen based on client’s feedback adapt
which ε-Greedy policy is used in selecting the best possible the video parameters. Now, theserver’s work can be
future action. outlined as: a) Video capture, b) Video transcoding, c)
The proposedalgorithms were implemented in accordance Video streaming, and d) Adapting video based on client’s
with the ITU-T J.247 recommendation (Table 1) which feedback.
describes about “objective perceptual video quality
measurement” [17]. 2.3. Client Side Functions

The client initializes the media player component to decode
2. PROPOSED SYSTEM
and play the video. It specifies the streaming URL as a
parameter to support media function, and connects with the
2.1. System Architecture
server using sockets by specifying the IP address and port
The architecture diagram of the proposed system is of server application. The client captures the packet using a
illustrated in Figure 1. It works on the top of HTTP in a packet capture framework. It calculates the throughput for
typical internet environment where the last mile certain period andassesses the frame rate (fps) and the bit

– 126 –