MILES (MachIne Learning for Efficient Streaming)

The project MILES (MachIne Learning for Efficient Streaming) was a 2019 winner of the PREMATURATION call for projects, led by the Innovation Lab of IP Paris.

Video streaming accounts for over half of internet video traffic (mostly DASH/HLS). A large share of this data is used over mobile/wireless connections. However, streaming to mobile users remains challenging due to the quickly changing network conditions. (For example: public transports users will move rapidly between different quality and type of coverage.)

For the last 10 years, the Multi-Media (MM) team of IDS/LTCI has developed GPAC, an open-source software dedicated to rich-media and broadcast technologies. This video streaming suite is recognized for its quality by leaders of the industry (received a Technology & Engineering Emmy® Awards in 2021, soon to be adopted by Netflix…)

MILES builds on the team expertise in adaptive data transfer to integrate Machine Learning based bitrate adaptation into video streaming.

• The MILES algorithm relies on a constrained online convex optimization approach for bitrate selection (reinforcement/online learning approach)

• MILES was first evaluated offline on a Wi-Fi testbed (preliminary throwaway prototype) and is now being live tested by users.

• MILES was fully integrated into the video streaming suite GPAC permitting its rapid deployment to the large user base.

• This Machine Learning toolbox is expected to give GPAC a strong competitive advantage especially in video streaming to mobile terminals.

• The teams developed a Python plugin backend to help the ML-application development. This was positively received by the community who exploited this new feature for unexpected new use-cases.