How Bilibili Turned Academic ABR Research into a Mobile Playback Boost

Background

Modern streaming services must support high‑definition video (1080p, 4K, 8K) while users often experience unstable downstream bandwidth, especially on mobile devices. Adaptive Bitrate (ABR) algorithms are used to select the appropriate resolution in real time to balance visual quality and smooth playback, thereby improving Quality of Experience (QoE).

Limitations of Traditional ABR Algorithms

Telemetry analysis identified three recurring failure scenarios:

When the network is fast, the initial low‑resolution start‑up chosen for fast playback frustrates users who manually switch to a higher resolution.

During poor network conditions, aggressive bitrate choices keep the resolution high, causing repeated stalls and prompting users to downgrade manually.

Users with strong data‑saving preferences deliberately select low resolutions, a behavior that vanilla ABR cannot address.

Root‑cause investigation showed that (a) start‑up resolution is fixed and cannot be adjusted by ABR, (b) seeking actions reset the buffer to near zero, and (c) prolonged low‑resolution periods after a temporary speed dip degrade user satisfaction.

Intelligent Resolution Solution

Real‑time QoE Model Based on Subjective Scores

Existing QoE formulas lack standardized weights, are linear, and evaluate only a single playback session. Bilibili conducted extensive subjective tests measuring start‑up time, resolution, resolution switches, stall count, and stall duration. The findings were:

Higher resolution yields higher scores, but with diminishing returns.

Users tolerate a single long stall better than many short stalls.

Start‑up stalls under three seconds have minimal impact compared with in‑playback stalls.

From these observations a memory‑effect‑aware QoE model was built. The model updates a QoE score continuously during playback:

Initial QoE is set by start‑up latency and start‑up resolution.

When a stall occurs, the score is penalized proportionally to recent stall frequency and duration.

Resolution switches cause no abrupt QoE jumps.

In the absence of stalls, the score converges toward a target value determined by the current resolution.

Network Speed Pre‑processing

Accurate bandwidth estimation is essential. Two systematic biases were identified:

Over‑estimation: At high download speeds, cached data makes the measured transfer time near zero, inflating the estimated speed. Mitigation: cap the maximum estimated speed.

Under‑estimation: CDN timeout‑retries add extra delay, reducing the estimated speed. Mitigation: detect server‑side timeout events and subtract the timeout duration from the measured transfer time.

Intelligent Resolution Framework

With refined bandwidth inputs and the new QoE model, a generalized adaptive resolution framework—named “Intelligent Resolution”—was constructed. Key components:

Long‑term network features: From three‑minute multi‑video traces, percentile‑based speed levels and volatility metrics are extracted. These features inform the aggressiveness of the decision‑making process.

Start‑up resolution module: Uses the recent 30‑second network statistics and long‑term features to select a higher initial resolution without increasing stall risk.

In‑playback resolution module: Retains the Pensieve neural‑network architecture but replaces the traditional QoE loss function with the memory‑aware real‑time QoE model. The retrained model yields roughly a 3 % average QoE improvement and a significant uplift of the minimum QoE.

Resolution up‑down decision center: Monitors network and buffer conditions to suppress excessive resolution changes and to dampen disturbances caused by user scrubbing or transient speed spikes.

Model Deployment on Mobile Clients

Neural‑network ABR models are computationally heavy for mobile devices. Following the 2019 “PiTree” approach, the trained Pensieve network is converted into a decision‑tree representation and then into native C++ code using the m2cgen library. The generated code consists of a few hundred lines, runs directly on the client, and incurs less than 1 % QoE loss compared with the original network.

User Preference Handling

Telemetry revealed three dominant user clusters:

HD: Prioritizes picture quality, tolerates stalls.

Auto: Seeks the highest feasible quality while keeping playback smooth.

Data‑saving: Minimizes bandwidth consumption regardless of visual quality.

The intelligent resolution algorithm is being extended to accept a user‑selected preference flag, allowing the system to adapt its aggressiveness and bandwidth budget accordingly.

Conclusion

By integrating academic ABR research, a memory‑aware QoE model, bias‑corrected bandwidth estimation, and lightweight client‑side deployment, the Intelligent Resolution system delivers a measurable QoE gain (≈3 % average improvement) with negligible client overhead. Future work includes studying latency effects on resolution switching and further audio‑video transport optimizations.