Deep Reinforcement Learning Based Routing in IP Media Broadcast Networks: Feasibility and Performance

Abstract

The Media Broadcast industry has evolved from Serial Digital Interface (SDI) based infrastructures to IP networks. While IP based video broadcast is well established in the data plane, the use of IP networks to transport media flows still poses challenges in terms of resource management and orchestration. SDN based orchestration architectures have emerged in the industry that use SDN to route the media flows of a broadcast service across the provider IP network. Several approaches to multimedia flow routing in IP based SDN networks have been proposed in the context of streaming applications over the Internet. These range from model based linear optimization solutions that have high complexity to simple shortest path based routing heuristics with either static or dynamic link costs. More recently modelfree optimization methods such as Deep Reinforcement Learning (DRL) have been proposed for routing and Traffic Engineering of multimedia flows in SDN networks. The media broadcast scenario however has specific requirements, with services like Master Control Room operation and Live broadcasting of events, and it has been rarely addressed in the literature.In this work we propose a DRL based routing method for this scenario and compare it to static and dynamic link cost algorithms based on Dijkstra shortest paths. This is to our knowledge the first work to follow this approach in the context of Media Broadcast services in IP infrastructures. The algorithm is designed considering the specifications and capabilities of one of the leading SDN orchestrators in the market and considers the more common Service Level Agreement requirements in the industry. Three different DRL algorithms are implemented and compared and we evaluate them using a real service provider network topology. The results indicate that DRL based routing is applicable in real production scenarios and that it achieves considerable performance gains when compared to the static and dynamic link cost shortest path algorithms commonly used today.

A Deep Reinforcement Learning-Based Optimal Transmission Control Method for Streaming Videos

“…Take t as the sampling start time, t+n.ΔT is the end time of the sample, Δt is the sampling interval, and equal interval sampling is performed on the physical layer transmission blocks of the wireless network. The specific calculation formula is as follows [30]: The optimization control method for streaming video transmission based on deep reinforcement learning can achieve better data transmission performance through realtime network state perception and active adjustment strategies based on user needs learning. Stability analysis can help us evaluate the robustness of optimized control methods.…”