| dc.description.abstract | The fundamental conflict between the increasing consumer demand for better Quality-of-Experience (QoE) and the limited supply of network resources has become significant challenges to modern video delivery systems.
State-of-the-art adaptive bitrate (ABR) streaming algorithms are dedicated to drain available bandwidth in hope to improve viewers' QoE, resulting in inefficient use of network resources.
In this thesis, we develop an alternative design paradigm, namely rate-distortion optimized streaming (RDOS), to balance the contrast demands from video consumers and service providers.
Distinct from the traditional bitrate maximization paradigm, RDOS must operate at any given point along the rate-distortion curve, as specified by a trade-off parameter.
The new paradigm has found plausible explanations in information theory, economics, and visual perception.
To instantiate the new philosophy, we decompose adaptive streaming algorithms into three mutually independent components, including throughput predictor, reward function, and bitrate selector.
We provide a unified framework to understand the connections among all existing ABR algorithms.
The new perspective also illustrates the fundamental limitations of each algorithm by going behind its underlying assumptions.
Based on the insights, we propose novel improvements to each of the three functional components.
To alleviate a series of unrealistic assumptions behind bitrate-based QoE models, we develop a theoretically-grounded objective QoE model.
The new objective QoE model combines the information from subject-rated streaming videos and the prior knowledge about human visual system (HVS) in a principled way.
By analyzing a corpus of psychophysical experiments, we show the QoE function estimation can be formulated as a projection onto convex sets problem.
The proposed model presents strong generalization capability over a broad range of source contents, video encoders, and viewing conditions.
Most importantly, the QoE model disentangles bitrate with quality, making it an ideal component in the RDOS framework.
In contrast to the existing throughput estimators that approximate the marginal probability distribution over all connections, we optimize the throughput predictor conditioned on each client.
Although there are lack of training data for each Internet Protocol connection, we can leverage the latest advances in meta learning to incorporate the knowledge embedded in similar tasks.
With a deliberately designed objective function, the algorithm learns to identify similar structures among different network characteristics from millions of realistic throughput traces.
During the test phase, the model can quickly adapt to connection-level network characteristics with only a small amount of training data from novel streaming video clients with a small number of gradient steps.
The enormous space of streaming videos, constantly progressing encoding schemes, and great diversity of throughput characteristics make it extremely challenging for modern data-driven bitrate selectors that are trained with limited samples to generalize well.
To this end, we propose a Bayesian bitrate selection algorithm by adaptively fusing an online, robust, and short-term optimal controller with an offline, susceptible, and long-term optimal planner.
Depending on the reliability of the two controllers in certain system states, the algorithm dynamically prioritizes the one of the two decision rules to obtain the optimal decision.
To faithfully evaluate the performance of RDOS, we construct a large-scale streaming video dataset -- the Waterloo Streaming Video database.
It contains a wide variety of high quality source contents, encoders, encoding profiles, realistic throughput traces, and viewing devices.
Extensive objective evaluation demonstrates the proposed algorithm can deliver identical QoE to state-of-the-art ABR algorithms at a much lower cost.
The improvement is also supported by so-far the largest subjective video quality assessment experiment. | en |
