Scholarlink

<p>This paper investigates combat strategies for swarms of agents with limited offensive, defensive, and communication capabilities within a confined area. To address the challenges of low convergence speed and inefficient information extraction in large-scale swarms, the actor-critic framework of the Deep Deterministic Policy Gradient (DDPG) algorithm is enhanced, incorporating a mean-field extraction module to aggregate local neighbor interactions and an attention extraction module to prioritize relevant environmental features. The components address two key limitations in large-scale multi-agent systems: the slow convergence rate caused by a variable number of agents and the constrained capacity to distill pertinent information from complex environments. Furthermore, a reward mechanism based on potential functions (PF-reward mechanism) is introduced to guide the strategy model in effectively balancing mean-field information with individual observations. Experimental results demonstrate that the proposed algorithm significantly improves convergence speed, learning efficiency, and stability. These benefits are amplified in larger-scale scenarios, with the final combat convergence rate increasing by <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\varvec{30\%-40\%}\)</EquationSource> </InlineEquation>. The refined network structure leverages information from both the individual agent and the mean-field extraction module, without requiring excessive attention to global information, avoiding the interference of low-importance information on strategy generation. At the same time, the PF-reward mechanism sharpens experience distinctions to better guide the iteration of parameters.</p>

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

An improved multi-agent combat algorithm for large-scale swarms based on DDPG

  • 期刊论文

  • 发表日期：

    2026年7月6日

查看全文
  • Yuxin Zhang,
  • Enjiao Zhao,
  • Dong Han,
  • Hong Liang

摘要

This paper investigates combat strategies for swarms of agents with limited offensive, defensive, and communication capabilities within a confined area. To address the challenges of low convergence speed and inefficient information extraction in large-scale swarms, the actor-critic framework of the Deep Deterministic Policy Gradient (DDPG) algorithm is enhanced, incorporating a mean-field extraction module to aggregate local neighbor interactions and an attention extraction module to prioritize relevant environmental features. The components address two key limitations in large-scale multi-agent systems: the slow convergence rate caused by a variable number of agents and the constrained capacity to distill pertinent information from complex environments. Furthermore, a reward mechanism based on potential functions (PF-reward mechanism) is introduced to guide the strategy model in effectively balancing mean-field information with individual observations. Experimental results demonstrate that the proposed algorithm significantly improves convergence speed, learning efficiency, and stability. These benefits are amplified in larger-scale scenarios, with the final combat convergence rate increasing by \(\varvec{30\%-40\%}\) . The refined network structure leverages information from both the individual agent and the mean-field extraction module, without requiring excessive attention to global information, avoiding the interference of low-importance information on strategy generation. At the same time, the PF-reward mechanism sharpens experience distinctions to better guide the iteration of parameters.