Video accounts for 56% of the current (2017) mobile network traffic and is expected to be responsible for 73% by 2023 . Consequently, optimization concerning video delivery has been gaining more and more interested and is the subject discussed in this article.
To adapt to the heterogeneity of networks and the variable available bandwidth, the state of the art in delivering video is the adaptive video streaming (ABR). The delivery of video with ABR is done by encoding the video in different quality levels, the higher the quality level, the more bandwidth is needed to deliver the video. To avoid video stalling (rebuffering, pause), which, in general, degrades the quality of experience (QoE) more than presenting the video in a lower quality, ABR works by estimating the maximum video quality that can be downloaded by the available network bandwidth.
Multiple problems can appear when ABR clients compete for resources, some examples are: the network is divided unfairly among clients, i.e., one client might get more bandwidth than the other; and instability in the selected video quality, i.e., the clients keep changing the video quality, which emphasize the video distortions and decrease QoE.
In the context of the SPOTLIGHT project, we are proposing new techniques to improve the overall QoE of ABR users in the presence of network congestion. To this end, we are applying recent machine learning techniques that can learn on experience and optimize the network, the so-called reinforcement learning.
The results of this research will allow the improved QoE for users without increasing infrastructure, i.e., clients will enjoy a higher QoE without increasing the bandwidth of their links. This is good from a financial perspective but is also good because it will allow more efficient use of resources.
 “Ericsson Mobility Report June 2018,” Ericsson, 2018.