Scholarlink

<p>Sparse rewards present a significant challenge in multi-agent reinforcement learning (MARL). This is due to limited feedback, which complicates the development of effective policies. In this work, we introduce <span>INSPIRE</span> (Individualized and Neighbor-based Sharing Prioritized Experience Replay), a framework designed to enhance training efficiency and performance under sparse reward conditions. INSPIRE decomposes team experience into individual experiences and restricts selective sharing to local neighborhoods. At its core, the Neighbor Experience Transmitter evaluates the value of experiences using neighbor feedback, enabling targeted exchange of high-value experiences. Empirical results on standard MARL benchmarks show that INSPIRE outperforms five state-of-the-art algorithms, achieving relative improvements of 13.60% on SMAC, 12.93% on SMACv2, and 37.92% on GRF. Ablation studies and comparative analyses further confirm its effectiveness across tasks and reward settings. These findings highlight the potential of INSPIRE for more efficient and generalizable training in multi-agent systems under sparse rewards.</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

Boosting Efficient Experience Exchange in Sparse-Reward Multi-Agent Reinforcement Learning

  • 期刊论文

  • 发表日期：

    2026年4月15日

查看全文
  • Jianan Zhang,
  • Zaipeng Xie,
  • Nuo Yang,
  • Juguang Jin,
  • WenZhan Song

摘要

Sparse rewards present a significant challenge in multi-agent reinforcement learning (MARL). This is due to limited feedback, which complicates the development of effective policies. In this work, we introduce INSPIRE (Individualized and Neighbor-based Sharing Prioritized Experience Replay), a framework designed to enhance training efficiency and performance under sparse reward conditions. INSPIRE decomposes team experience into individual experiences and restricts selective sharing to local neighborhoods. At its core, the Neighbor Experience Transmitter evaluates the value of experiences using neighbor feedback, enabling targeted exchange of high-value experiences. Empirical results on standard MARL benchmarks show that INSPIRE outperforms five state-of-the-art algorithms, achieving relative improvements of 13.60% on SMAC, 12.93% on SMACv2, and 37.92% on GRF. Ablation studies and comparative analyses further confirm its effectiveness across tasks and reward settings. These findings highlight the potential of INSPIRE for more efficient and generalizable training in multi-agent systems under sparse rewards.