跳到正文
arXiv cs.AI· Jiwon Jeon, Myungsik Cho, Woojun Kim, Seongmin Kim, Woohyeon Byeon, Seungyul Han, Youngchul Sung·· 4 小时前AI 评分29

MASTARS:基于回报条件子目标的多智能体序列轨迹增强框架

MASTARS: Multi-Agent Sequential Trajectory Augmentation with Return-Conditioned Subgoals

AI 导读

MASTARS 是一种扩散式多智能体轨迹增强框架,针对离线多智能体强化学习中独立生成缺乏协调、联合生成缺乏多样性的两难,采用逐智能体序列生成并借助 inpainting 保持跨智能体协调一致。实验在 MPE、SMAC、SMAC-v2 等基准上显著提升多种离线 MARL 方法的性能。

正文 · 原文

View PDF HTML (experimental)

Abstract:The performance of offline Multi-Agent Reinforcement Learning (MARL) is limited by the quality of the fixed offline dataset, motivating trajectory augmentation. However, augmenting multi-agent trajectories presents a fundamental challenge of the diversity-coordination trade-off: independent generation improves diversity but lacks coordination, while joint generation induces coordination but lacks diversity, reproducing the training distribution. To address this challenge in multi-agent data augmentation, we introduce MASTARS, a novel diffusion-based framework that generates multi-agent trajectories sequentially across agents while enforcing cross-agent consistent coordination via the method of inpainting. MASTARS produces trajectories that are diverse, coordinated and globally coherent. Experiments on various benchmarks such as MPE, SMAC, and SMAC-v2 demonstrate significant performance gains across multiple offline MARL methods.
Subjects: Artificial Intelligence (cs.AI)
Cite as: arXiv:2611.17793 [cs.AI]
  (or arXiv:2611.17793v1 [cs.AI] for this version)
  https://doi.org/10.48550/arXiv.2611.17793

arXiv-issued DOI via DataCite (pending registration)

Submission history

From: Jiwon Jeon [view email]
[v1] Fri, 13 Nov 2026 06:12:14 UTC (1,393 KB)

来源:arXiv cs.AI · arxiv.org