[Paper Review] Optimal Decentralized Dynamic Policies for Video Streaming over Wireless Channels

Rahul Singh, P. R. Kumar|arXiv (Cornell University)|Feb 20, 2019

Image and Video Quality Assessment25 references4 citations

TL;DR

This paper proposes a decentralized, dynamic policy for optimizing video streaming quality of experience (QoE) over wireless channels by enabling clients to independently select video resolution and transmission power based on a globally announced energy price. The key contribution is a provably optimal, distributed solution using duality-based pricing and index policies, with simulations showing significant QoE gains over baseline methods like Round Robin and Shortest Queue.

ABSTRACT

The problem addressed is that of optimally controlling, in a decentralized fashion, the download of mobile video, which is expected to comprise 75 % of total mobile data traffic by 2020. The server can dynamically choose which packets to download to clients, from among several packets which encode their videos at different resolutions, as well as the power levels of their transmissions. This allows it to control packet delivery probabilities, and thereby, for example, avert imminent video outages at clients. It must however respect the access point's constraints on bandwidth and average transmission power. The goal is to maximize video "Quality of Experience" (QoE), which depends on several factors such as (i) outage duration when the video playback buffer is empty, (ii) number of outage periods, (iii) how many frames downloaded are of lower resolution, (iv) temporal variations in resolution, etc. It is shown that there exists an optimal decentralized solution where the AP announces the price of energy, and each client distributedly and dynamically maximizes its own QoE subject to the cost of energy. A distributed iterative algorithm to solve for optimal decentralized policy is also presented. Further, for the client-level QoE optimization, the optimal choice of video-resolution and power-level of packet transmissions has a simple monotonicity and threshold structure vis-a-vis video playback buffer level. When the number of orthogonal channels is less than the number of clients, there is an index policy for prioritizing packet transmissions. When the AP has to simply choose which clients' packets to transmit, the index policy is asymptotically optimal as the number of channels is scaled up with clients.

Motivation & Objective

• To address the challenge of maximizing Quality of Experience (QoE) in video streaming over unreliable wireless networks with dynamic channel conditions and client buffer states.
• To design a decentralized control policy where clients make individual decisions on resolution and transmission power based on local buffer states and a global energy price.
• To ensure optimal QoE under aggregate power and bandwidth constraints at the access point (AP), while maintaining computational tractability.
• To develop index-based scheduling policies when the number of orthogonal channels is less than the number of clients.
• To evaluate the performance of proposed policies against standard benchmarks like Round Robin and Shortest Queue in dynamic wireless environments.

Proposed method

• The AP announces a dual variable (energy price) λ⋆ that equilibrates total energy consumption with available power, enabling clients to optimize their individual QoE under cost constraints.
• A distributed iterative algorithm computes the optimal energy price λ⋆ by solving a linear program with variables scaling linearly with the number of clients.
• For single-client optimization, the optimal resolution and power policy exhibits a monotonic threshold structure based on buffer level and channel quality.
• When M < N orthogonal channels are available, a one-step look-ahead index policy is derived using the multi-armed bandit superprocess framework.
• The index policy is constructed by improving a base policy (e.g., Round Robin) via the look-ahead rule, ensuring asymptotic optimality as the number of channels scales with clients.
• The solution leverages duality theory to prove that the decentralized policy is optimal for the constrained Markov decision process (CMDP) formulation.

Experimental results

Research questions

• RQ1 Can a decentralized policy be designed such that each client independently optimizes its QoE while respecting global power and bandwidth constraints?
• RQ2 What is the optimal structure of the resolution and transmission power policy for a single client under varying buffer levels and channel conditions?
• RQ3 How can the AP efficiently schedule multiple clients when the number of available channels is limited?
• RQ4 How does the performance of the proposed index policy compare to traditional scheduling policies like Round Robin and Shortest Queue in dynamic wireless environments?
• RQ5 Under what conditions is the proposed decentralized policy provably optimal for maximizing cumulative QoE?

Key findings

• The optimal decentralized policy is achieved when the AP announces a price λ⋆ per unit energy, and each client independently maximizes its QoE subject to this cost, resulting in a globally optimal solution.
• The optimal policy for a single client exhibits a monotonic threshold structure: higher buffer levels favor higher resolution and lower power, while low buffers trigger lower resolution and higher power for reliability.
• When M < N orthogonal channels are available, the proposed look-ahead index policy significantly outperforms Round Robin and Shortest Queue policies in QoE, especially under high channel variability.
• The energy price λ⋆ decreases with increasing available power and increases with the number of clients, reflecting the trade-off between resource availability and demand.
• As the number of channels scales with the number of clients, the index policy becomes asymptotically optimal, confirming its scalability and robustness.
• Simulations show that the proposed policy achieves higher cumulative QoE than baseline methods, with gains up to 20% in certain configurations, particularly when channel reliability varies.