Technical Deep Dive of Adaptive Streaming for “The Longest Day in Chang'an”: AI, Big Data, and QoE Optimization

Technical experts from Alibaba Entertainment present the core technologies behind the streaming experience of "The Longest Day in Chang'an".

The talk covers the "Smart Bitrate" (智能档) system that combines big‑data analysis and AI‑driven machine learning to dynamically select the optimal video quality (SD, HD, or Blu‑ray) based on network conditions and user preferences.

Why buffering occurs

Buffering happens when the video player enters a "hungry" state due to insufficient video data. Real‑world factors such as poor network signal, high bitrate peaks, and bandwidth competition exacerbate this problem.

Smart Bitrate principle

Video files of various bitrates and resolutions are split into 5‑10 second segments. An algorithm makes real‑time decisions on which segment resolution to serve, dynamically raising or lowering quality to match fluctuating network conditions and avoid buffering.

Challenges

1. Balancing high definition and smooth playback, often modeled as a QoE (Quality of Experience) optimization problem that weighs playback duration, stall count, and resolution switches.

2. Bridging the gap between offline algorithm training and online performance, addressed by using MPC (Model Predictive Control) and deep‑learning approaches such as the Pensieve algorithm.

3. Delivering a truly "intelligent" user experience that respects user preferences (e.g., preferring high‑definition when possible) rather than defaulting to overly conservative low‑quality streams.

Technical solutions

The QoE formula is normalized over time and supplemented with intermediate metrics (e.g., bitrate switch frequency, buffer events, watch time) for fine‑grained monitoring and offline optimization.

Multiple algorithmic strategies are employed:

Rule‑based and buffer‑based heuristics.

MPC‑based control that ingests bandwidth, bitrate, and buffer length to select the optimal playback plan.

Deep‑learning models (Pensieve) trained on large‑scale user behavior data, with ongoing efforts to improve generalization.

High‑quality data collection is emphasized as essential for model training and continuous improvement.

Results

Compared to industry baselines, the proprietary solution achieves a 5‑10% increase in high‑definition playback duration and a 10‑20% reduction in stall events. User‑specific preferences are satisfied: users favoring high‑definition enjoy optimal quality for 92% of their watch time, while bandwidth‑constrained users receive smooth playback at appropriate resolutions.

In summary, the combination of big‑data analytics, AI/ML algorithms, and QoE‑driven optimization enables a near‑zero‑stall streaming experience for "The Longest Day in Chang'an".