Tingyao Wu

Design and optimisation of a FAQ-learning-based HTTP adaptive streaming client

In recent years, HTTP (Hypertext Transfer Protocol) adaptive streaming (HAS) has become the de facto standard for adaptive video streaming services. A HAS video consists of multiple segments, encoded at multiple quality levels. State-of-the-art HAS clients employ deterministic heuristics to dynamically adapt the requested quality level based on the perceived network conditions. Current HAS client heuristics are, however, hardwired to fit specific network configurations, making them less flexible to fit a vast range of settings. In this article, a (frequency adjusted) Q-learning HAS client is proposed. In contrast to existing heuristics, the proposed HAS client dynamically learns the optimal behaviour corresponding to the current network environment in order to optimise the quality of experience. Furthermore, the client has been optimised both in terms of global performance and convergence speed. Thorough evaluations show that the proposed client can outperform deterministic algorithms by 11–18% in terms of mean opinion score in a wide range of network configurations.

SVC-based HTTP adaptive streaming

HTTP adaptive streaming (HAS) is rapidly evolving into a key video delivery technology, supported by implementations from Microsoft, Apple, and Adobe, and actively pursued by standardization organizations. Using segments in multiple video qualities, distributed via an already available Hypertext Transfer Protocol (HTTP) delivery infrastructure, a HAS client is able to seamlessly adapt to the available bandwidth in the network. However, existing HAS solutions have a number of disadvantages such as the additional storage and bandwidth requirements, a large playout buffer to absorb network impairments, and a non-optimal quality selection under fluctuating network conditions. In this paper, we investigate the opportunity of combining HAS with scalable video coding. We show that this combination creates possibilities to reduce the client buffer, which implies improvements for live and interactive video, and reduces storage requirements, increases the cache hit-ratio for supporting content delivery network (CDN) nodes, and demonstrates more robust behavior in the HAS client, ultimately improving the quality of experience (QoE) for the viewer.

On the Use of Reservoir Computing in Popularity Prediction

Predicting the life cycle and the short-term popularity of a Web object is important for network architecture optimization. In this paper, we attempt to predict the popularity of a Web object given its historical access records using a novel neural network technique, reservoir computing (RC). The traces of popular videos at YouTube for five continuous months are taken as a case study. We compare RC with existing analytical models. Experimental results show that RC, given a 10-day trace composed of daily cumulative views for a video, is able to predict the next-day’s popularity with less than 5% relative square errors (RSEs). It is also demonstrated that RC achieves the best prediction performance among all compared models in longer-term prediction. The advantages and limitations of using RC in popularity prediction are discussed.

Feature subset selection for improved native accent identification

In this paper, we develop methods to identify accents of native speakers. Accent identification differs from other speaker classification tasks because accents may differ in a limited number of phonemes only and moreover the differences can be quite subtle. In this paper, it is shown that in such cases it is essential to select a small subset of discriminative features that can be reliably estimated and at the same time discard non-discriminative and noisy features. For identification purposes a speaker is modeled by a supervector containing the mean values for the features for all phonemes. Initial accent models are obtained as class means from the speaker supervectors. Then feature subset selection is performed by applying either ANOVA (analysis of variance), LDA (linear discriminant analysis), SVM-RFE (support vector machine-recursive feature elimination), or their hybrids, resulting in a reduced dimensionality of the speaker vector and more importantly a significantly enhanced recognition performance. We also compare the performance of GMM, LDA and SVM as classifiers on a full or a reduced feature subset. The methods are tested on a Flemish read speech database with speakers classified in five regions. The difficulty of the task is confirmed by a human listening experiment. We show that a relative improvement of more than 20% in accent recognition rate can be achieved with feature subset selection irrespective of the choice of classifier. We finally show that the construction of speaker-based supervectors significantly enhances results over a reference GMM system that uses the raw feature vectors directly as input, both in text dependent and independent conditions.

Session reconstruction for HTTP adaptive streaming: laying the foundation for network-based QoE monitoring

HTTP Adaptive Streaming (HAS) is rapidly becoming a key video delivery technology for fixed and mobile networks. However, today there is no solution that allows network operators or CDN providers to perform network-based QoE monitoring for HAS sessions. We present a HAS QoE monitoring system, based on data collected in the network, without monitoring information from the client. To retrieve the major QoE parameters such as average quality, quality variation, rebuffering events and interactivity delay, we propose a technique called session reconstruction. We define a number of iterative steps and developed algorithms that can be used to perform HAS session reconstruction. Finally, we present the results of a working prototype for the reconstruction and monitoring of Microsoft Smooth Streaming HAS sessions that is capable of dealing with intermediate caching and user interactivity. We describe the main observations when using the platform to analyze more than a hundred HAS sessions.

Reuse time based caching policy for video streaming

In this paper, we investigate the proxy video caching problem and propose a specific caching replacement strategy for segmented video streaming. The proposed caching algorithm makes use of the natural linear time structure of video streaming, which fits the philosophy of the optimal MIN algorithm. By observing the watching positions of active viewers at a certain point of time, the replacement approach computes the exact reuse time or, if not available, the predicted reuse time for stored video segments, and discards the segment that is supposedly to be used the furthest in time. Evaluations performed on a trace of a real Video-On-Demand (VOD) service show that the proposed caching algorithm substantially improves the hit ratio, especially for small caches.

Scalable network-based video-freeze detection for HTTP adaptive streaming

HTTP adaptive streaming (HAS) has become a key video delivery technology for mobile and fixed networks. Internet service providers and CDN (Content Delivery Network) providers are interested in network-based monitoring the clients Quality of Experience (QoE) for HAS sessions. In our previous work, we designed a HAS QoE monitoring system based on the sequence of HTTP GET requests collected at the CDN nodes. The system relies on a technique called session reconstruction to retrieve the major QoE parameters without modification of the clients. However, session reconstruction is computationally intensive and requires manual configuration of reconstruction rules. To overcome the limitations of session reconstruction, this paper proposes a scalable machine learning (ML) based scheme that detects video freezes using a few high-level features extracted from the network-based monitoring data. We determine the discriminative features for session representation and assess five potential classifiers. We select the C4.5 decision tree as classifier because of its simplicity, scalability, accuracy, and explainability. To evaluate our solution, we use traces of Apple HTTP Live Streaming video sessions obtained from a number of operational CDN nodes and traces of Microsoft Smooth Streaming video sessions acquired in a controlled lab environment. Experimental results show that an accuracy of about 98%, 98%, and 90% can be obtained for the detection of a video freeze, a long video freeze, and multiple video freezes, respectively. Excluding log parsing, the computational cost of the proposed video-freeze detection is 33 times smaller than needed for session reconstruction.

Maximum Likelihood Based Temporal Frame Selection

In this paper, we propose a maximum likelihood (ML) based frame selection approach. A fixed frame rate adopted in most state-of-the-art speech recognition systems can face some problems, such as accidentally meeting noisy frames, assigning the same importance to each frame, and pitch asynchronous representation. As an attempt to avoid those problems, our approach selects reliable frames from a fine resolution along the time axis in a phoneme recognition task, we show that significant improvements are achieved with the frame selection approach comparing to a system with a fixed frame rate

Factor Selection for Reinforcement Learning in HTTP Adaptive Streaming

At present, HTTP Adaptive Streaming (HAS) is developing into a key technology for video delivery over the Internet. In this delivery strategy, the client proactively and adaptively requests a quality version of chunked video segments based on its playback buffer, the perceived network bandwidth and other relevant factors. In this paper, we discuss the use of reinforcement-learning (RL) to learn the optimal request strategy at the HAS client by progressively maximizing a pre-defined Quality of Experience (QoE)-related reward function. Under the framework of RL, we investigate the most influential factors for the request strategy, using a forward variable selection algorithm. The performance of the RL-based HAS client is evaluated by a Video-on-Demand (VOD) simulation system. Results show that given the QoE-related reward function, the RL-based HAS client is able to optimize the quantitative QoE. Comparing with a conventional HAS system, the RL-based HAS client is more robust and flexible under versatile network conditions.

Network-based video freeze detection and prediction in HTTP adaptive streaming

Given the popularity of HTTP adaptive streaming (HAS) technology for media delivery over mobile and fixed networks, the clients Quality of Experience (QoE) for HAS video sessions is of particular interest for network providers and Content Delivery Network (CDN) operators. Despite that, network providers are not able to directly obtain QoE-relevant metrics such as video freezes, initial buffering time, and the frequency of quality switches from the client. This paper proposes a scalable machine learning (ML) based scheme that offline detects and online predicts video freezes using a few features extracted from the network-based monitoring data, i.e., a sequence of HTTP GET requests sent from the video client. We determine the discriminative features for detecting video freezes based on multi-scale windows using the criterion of information gain (IG). Four traditional classifiers are investigated and the C4.5 decision tree is selected because of its simplicity, scalability, accuracy, and interpretability. Our approach for session-based offline freeze detection is evaluated on the Apple HTTP Live Streaming video sessions obtained from a number of operational CDN nodes and on the traces of Microsoft Smooth Streaming video sessions acquired in a controlled lab environment. Experimental results show that, even with the disturbance of user interactivity, an accuracy of about 91% can be obtained for the detection of a video freeze, a long video freeze, and multiple video freezes. The experiments for the online freeze prediction show that more than 30% of the video freezes can be foreseen one segment in advance.