Skip to content
Open access

An efficient on-policy actor–critic architecture for adaptive bitrate streaming using deep reinforcement learning

Sep 2026 · Scientific Reports · 0 citations

TL;DR

The proposed GRASP algorithm outperforms Comyco and other cutting-edge DRL-based ABR methods under a variety of network situations and reduces the stall ratio compared to Comyco up to 67% and 70% for 5G and 4G networks respectively.

Abstract

Adaptive video streaming is a widely used way of distributing video material over the Internet. It employs adaptive bitrate (ABR) algorithms to ensure a high quality of experience (QoE) for end users under a variety of network situations. However, the dynamic and varied characteristics of emerging 5th generation (5G) and beyond 5G (B5G) networks frequently causes instability in ABR algorithms. Despite advances in learning-based ABR algorithms, particularly those that use cutting-edge deep reinforcement learning (DRL) technology, they typically achieve mediocre performance and struggle to offer optimal results consistently across a variety of network environments. Recent observations reveal that Comyco, a quality-aware adaptive video streaming technology, provides great QoE over 3G networks but suffers limits across 4G and 5G networks. To solve this issue, we offer GRAdient Sharing actor-critic approach with Policy optimization (GRASP), a revolutionary on-policy DRL-based ABR technique. We compare GRASP to various on-policy and off-policy DRL techniques such as Pensieve, GRAAB, Comyco, PPO-ABR and SAC-ABR and other cutting-edge DRL-based ABR algorithms and non DRL-based ABR algorithms including BB, RB and BOLA. For comparison, we utilize mm Wave 4G traces and Lumos 5G dataset. Our results show that the proposed GRASP algorithm outperforms Comyco and other cutting-edge DRL-based ABR methods under a variety of network situations. GRASP consistently achieves higher QoE over the Comyco DRL-based framework up to 11.44%, and 16.56% for 5G and 4G networks respectively, and even higher QoE for other DRL-based and non DRL-based ABR schemes. Meanwhile, GRASP reduces the stall ratio compared to Comyco up to 67% and 70% for 5G and 4G networks respectively.

Read PDF

Similar papers

Sep 2026

ESM: energy-efficient bitrate adaptation for video streaming with deep reinforcement learning

Experimental results demonstrate that the proposed ESM algorithm achieves well-balanced multi-objective performance for mobile video streaming under hardware resource constraints, and the achieved optimization gains will fluctuate slightly across different combinations of video content and network conditions.

Wen-Long Ke, Ying Wang, Yong Zhang · 0 citations

Towards Better Training Signal: Advantage Clipped Policy Optimization

ACPO is a practical and effective alternative to conventional IS-ratio clipping for RL post-training of LLMs, and establishes a connection between ACPO and gradient clipping in policy mirror descent (PMD), which is a standard technique to stabilize optimization process.

Rui-Chuan Huang, Jing-Han Liu, Cong-Liang Chen · 0 citations

Experimental Research of Hybrid Reinforcement Learning Methods for QoS-Aware Web Service Composition: an Actor-Critic Approach

A hybrid framework that integrates Reinforcement Learning with rule-based constraints, including action masking, no-repetition policies, and SLA-based limitations is proposed, which enables the reduction of the search space and improves the validity of generated service compositions.

I. Grishanova, J. Rogushina, Pylyp Andon · 0 citations
Open access Oct 2026

DIA: A Deep Reinforcement Learning-Based Interest-Aware ABR Scheme for Cloud Gaming

A Deep Reinforcement Learning-based Interest-Aware Adaptive Bitrate (ABR) scheme, called DIA, that guarantees the quality of perceptually key objects in each frame, while dynamically allocating lower resolutions for non-key objects to conserve bandwidth under constrained network conditions is proposed.

Zhe Zhang, Haiwei Dong, Marc St-Hilaire et al. · 0 citations

Learn from the Gap: Differential-Aware Advantage Pruning with Adaptive Rollout Sampling for GRPO

This work proposes FastRL, a novel plug-and-play reinforcement learning framework that simultaneously improves training efficiency and the effectiveness of policy learning, and introduces an advantage-aware pruning strategy to selectively preserve high-advantage trajectories while maximizing inter-trajectory gradient d...

Jia-Hua Yang, Zhiwei Yang, Xian-Peng Zhang et al. · 0 citations

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.