CBA Contextual Quality Adaptation for Adaptive Bitrate Video
CBA: Contextual Quality Adaptation for Adaptive Bitrate Video Streaming by Bastian Alt, Trevor Ballard, Ralf Steinmetz, Heinz Koeppl, Amr Rizk In Infocom 2019 Presented by: Hongpeng Guo Sep 17, 2019
Method Overview. • When using ABR algorithms to determine video quality online, we may rely on (1) Information gathered from client side. (measured throughput, buffer filling, and derivatives thereof ) i. e. DASH architecture. (2) Server or network assistance. (accurate network throughput measurement, source recommendations) i. e. SAND architecture. • The fundamental problem here is to determine how valuable either information is for the adaptation decision
Method Overview. • Sequential decision-making under uncertainty • Using network assistance and client-side information. (high dimensional network context) • Model the decision on a video segment quality as a contextual multi-armed bandit problem • Aiming to optimize an objective Qo. E : (i) the average video quality bitrate, (ii) the quality degradation, (iii) the video stalling.
Background – Multi-armed bandit problem. • A gambler at a row of slot machines, who has to decide (1) which machines to play, (2) how many times to play each machine and (3) in which order to play them, and (4) whether to continue with the current machine or try a different machine. • A classic online reinforcement learning problem that exemplifies the exploration-exploitation trade off.
Background – Multi-armed bandit problem Formation.
Background – Contextual Multi-armed bandit problem. • In each iteration an agent has to choose between arms. • Before making the choice, the agent sees a d-dimensional feature vector (context vector), associated with the current iteration. • The learner uses these context vectors along with the rewards of the arms played in the past to make the choice of the arm to play in the current iteration. • Over time, the learner's aim is to collect enough information about how the context vectors and rewards relate to each other, so that it can predict the next best arm to play by looking at the feature vectors
ABR Contextual Multi-armed bandit problem.
ABR Contextual Multi-armed bandit problem.
CBA-UCB The Contextual Bayes-UCB Algorithm.
CBA Other Sub-modules. • After each step, update the posterior distribution. • In order to achieve more fast computation, several sub-modules are introduced together with the CBA-UCB algorithms to reduce vector dimension while updating the posterior distribution. These submodules are o Variational Bayesian Inference (VB) o Stochastic Variational Inference (SVI) o One Step Stochastic Variational Inference (OS-SVI)
Simulations. • Test the regrets of CBA algorithms comparing with other bandit algorithm baselines.
Experiments. • Experiment on Name Data Network (NDN), which contains adequate network context features. o The contextual vector is of 101 dimension. o A Doubles NDN Topology as follows.
Experiment Results. • Comparing with Line. UCB baseline and two state-of-the art ABR algorithm: PANDA and BOLA
Pros and Cons. • Pros: o Focus on heterogenous network-assisted data for Video ABR, the model is more powerful than classic ABR algorithms who only play with client measurements. o Using online learning method to learn the important metrics. The algorithm is efficient and outperforms state-of-the-art ABR algorithms on Qo. E fairness. • Cons: o Experiment setting is simple, a testbed on larger networks is needed. o The CBA algorithms did not show much advantages over the baselines on cumulative Qo. E.
- Slides: 14