Scholarlink

This paper presents a deep reinforcement learning (DRL)–based congestion control framework enhanced by a Least Squares Congestion Minimization (LSCM) module to improve latency, bandwidth utilization, packet loss, and fairness under dynamic and heterogeneous network conditions. Unlike conventional TCP variants (e.g., Cubic, BBR) and existing DRL-based approaches that rely solely on instantaneous congestion signals, the proposed framework integrates residual-based congestion trend estimation into an Actor–Critic learning architecture. The LSCM module analytically estimates queue evolution residuals over a sliding observation horizon and embeds this information into multi-objective reward shaping, thereby reducing learning variance and stabilizing policy convergence. Extensive NS-3 simulations under mixed IoT, video streaming, and cloud-gaming traffic demonstrate up to a 24% reduction in RTT, a 19% improvement in bottleneck utilization, a 30% reduction in packet loss, and consistently high fairness (Jain’s index >0.94) compared to TCP Cubic, TCP BBR, and a DRL-only baseline. The results confirm that residual-guided learning provides a robust and scalable solution for intelligent congestion control in modern packet-switched networks.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Deep Reinforcement Learning Framework with Least Squares Congestion Minimization for Intelligent Network Congestion Control and Latency–Bandwidth Optimization

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Van Thanh Nguyen,
  • Kim Quoc Nguyen,
  • Ducthinh Nguyen,
  • Van Cuu Ho,
  • Leminhthien Huynh

摘要

This paper presents a deep reinforcement learning (DRL)–based congestion control framework enhanced by a Least Squares Congestion Minimization (LSCM) module to improve latency, bandwidth utilization, packet loss, and fairness under dynamic and heterogeneous network conditions. Unlike conventional TCP variants (e.g., Cubic, BBR) and existing DRL-based approaches that rely solely on instantaneous congestion signals, the proposed framework integrates residual-based congestion trend estimation into an Actor–Critic learning architecture. The LSCM module analytically estimates queue evolution residuals over a sliding observation horizon and embeds this information into multi-objective reward shaping, thereby reducing learning variance and stabilizing policy convergence. Extensive NS-3 simulations under mixed IoT, video streaming, and cloud-gaming traffic demonstrate up to a 24% reduction in RTT, a 19% improvement in bottleneck utilization, a 30% reduction in packet loss, and consistently high fairness (Jain’s index >0.94) compared to TCP Cubic, TCP BBR, and a DRL-only baseline. The results confirm that residual-guided learning provides a robust and scalable solution for intelligent congestion control in modern packet-switched networks.