Answers are grounded in this paper and cite their sources.
HTTP Adaptive Streaming (HAS) typically adapts video bitrate while using a fixed chunk duration. However, short chunks provide faster adaptation at the cost of more HTTP requests, whereas long chunks reduce request frequency but respond more slowly to network changes. This paper proposes Safety-Projected Factorized ACTBR, an application-layer framework that adapts both video quality and logical chunk duration. The video is divided into aligned 2-s Dynamic Adaptive Streaming over HTTP (DASH) media units, which can be combined to form logical requests of 2, 4, 6, or 10 s. A source bitrate-only Deep Deterministic Policy Gradient (DDPG) controller is first trained using fixed 2-s requests. Its quality actor is then transferred and frozen, while a separate duration actor is trained. A conservative safety projection limits unsafe duration choices before each request is executed. The simulator uses exact variable-bit-rate media-unit sizes and measured mobile bandwidth traces. The proposed framework is evaluated using 40 mobile bandwidth traces, three DASH videos, fixed training–validation–test splits, and five independent training seeds. Compared with the source Adaptive Bitrate (ABR)-only controller using fixed 2-s requests, ACTBR reduces the number of HTTP requests by 6.55% while changing the mean time-weighted nominal representation bitrate by only $$-\,0.0099$$ Mbps. This request reduction is accompanied by a 0.8528-point decrease in mean quality of experience (QoE) and a 0.1341-s increase in mean rebuffering. Fixed-duration experiments show that D4, D6, and D10 reduce requests more strongly but cause progressively larger QoE losses. Additional variable-bit-rate (VBR)–constant-bit-rate (CBR), shared-bottleneck, and runtime analyses show that the observed trade-off remains consistent and that central processing unit (CPU) inference is sub-millisecond. Overall, ACTBR provides a conservative balance between streaming QoE and HTTP request efficiency.
No takes yet. Share an insight, caveat, or question.
Chatri et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: