Scholarlink

Convolutional Neural Networks (CNNs) have achieved remarkable success in various domains, including computer vision, natural language processing, speech recognition, and autonomous driving. However, as task complexity and performance demands increase, there is a need for an exponential growth in network parameters, which intensifies the challenges of deploying such models on resource-constrained edge devices. Model compression has advanced substantially in addressing these limitations. Nevertheless, most existing approaches rely on single-method strategies and neglect the complex coupling relationships within network architectures, while often requiring extensive manual intervention during the compression process. To this end, this paper proposes a model compression approach that jointly optimizes pruning and quantization using Multi-Agent Reinforcement Learning. The network is partitioned into groups based on dependency, with an agent assigned to each group. These agents collaboratively coordinate their actions to maximize compression while preserving critical architectural. When executing compression actions, each agent accounts for the actions of others, and rewards are allocated through an explicit credit assignment mechanism based on individual contributions. We conducted experiments on representative architectures, including ResNet, VGG, and DenseNet, to evaluate the effectiveness of our method. The results demonstrate that our approach achieves satisfactory performance by significantly reducing both FLOPs and the number of parameters while maintaining high model accuracy.

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

HybCompress-MARL: Hybrid Model Compression with Multi-agent Reinforcement Learning for CNN

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Qianxi Li,
  • Yuze Huang,
  • Wenhui Zhang,
  • Wenchuan Xiong,
  • Ning Shi

摘要

Convolutional Neural Networks (CNNs) have achieved remarkable success in various domains, including computer vision, natural language processing, speech recognition, and autonomous driving. However, as task complexity and performance demands increase, there is a need for an exponential growth in network parameters, which intensifies the challenges of deploying such models on resource-constrained edge devices. Model compression has advanced substantially in addressing these limitations. Nevertheless, most existing approaches rely on single-method strategies and neglect the complex coupling relationships within network architectures, while often requiring extensive manual intervention during the compression process. To this end, this paper proposes a model compression approach that jointly optimizes pruning and quantization using Multi-Agent Reinforcement Learning. The network is partitioned into groups based on dependency, with an agent assigned to each group. These agents collaboratively coordinate their actions to maximize compression while preserving critical architectural. When executing compression actions, each agent accounts for the actions of others, and rewards are allocated through an explicit credit assignment mechanism based on individual contributions. We conducted experiments on representative architectures, including ResNet, VGG, and DenseNet, to evaluate the effectiveness of our method. The results demonstrate that our approach achieves satisfactory performance by significantly reducing both FLOPs and the number of parameters while maintaining high model accuracy.