arXivDaily arXiv每日学术速递 周一至周五更新
北美退伍军人节(Veterans Day),12 日无更新
arXiv 2602.06599cs.MAcs.AIcs.LG
收藏

基于联合经验最佳响应的高效策略空间响应 oracle

Sample-Efficient Policy Space Response Oracles with Joint Experience Best Response

Utrecht University(乌得勒支大学);Delft University of Technology(代尔夫特理工大学);Eindhoven University of Technology(埃因霍温理工大学)

Ariyan Bighashdel, Thiago D. Simão, Frans A. Oliehoek

AI总结:

本文提出联合经验最佳响应方法,通过离线强化学习提升 PSRO 的样本效率和实用性,实现高效多智能体策略学习。

摘要:

多智能体强化学习(MARL)提供了一种可扩展的替代方案,以替代精确博弈论分析,但其面临非平稳性和需要维持多样化的策略群体以捕捉非传递性互动的问题。策略空间响应 oracle(PSRO)通过迭代扩展受限游戏并使用近似最佳响应(BRs)来解决这些问题,但单个智能体的 BR 训练在多智能体或模拟成本较高的情况下会变得过于昂贵。我们引入了联合经验最佳响应(JBR),这是一种对 PSRO 的改进方法,它在当前元策略配置下收集轨迹,并重用这个联合数据集来同时计算所有智能体的 BRs。这消除了环境交互的成本并提高了最佳响应计算的样本效率。因为 JBR 将 BR 计算转换为离线强化学习问题,我们提出了三种缓解分布偏移偏差的措施:(i)保守 JBR 通过安全策略改进,(ii)探索增强的 JBR 通过扰动数据收集并允许理论保证,(iii)混合 BR 在 JBR 和周期性独立 BR 更新之间交替。在基准多智能体环境中,探索增强的 JBR 实现了最佳的准确率-效率权衡,而混合 BR 在样本成本仅为 fraction 的情况下实现了接近 PSRO 的性能。总体而言,JBR 使 PSRO 在大规模战略学习中更加实用,同时保持了均衡的鲁棒性。

英文摘要:

Multi-agent reinforcement learning (MARL) offers a scalable alternative to exact game-theoretic analysis but suffers from non-stationarity and the need to maintain diverse populations of strategies that capture non-transitive interactions. Policy Space Response Oracles (PSRO) address these issues by iteratively expanding a restricted game with approximate best responses (BRs), yet per-agent BR training makes it prohibitively expensive in many-agent or simulator-expensive settings. We introduce Joint Experience Best Response (JBR), a drop-in modification to PSRO that collects trajectories once under the current meta-strategy profile and reuses this joint dataset to compute BRs for all agents simultaneously. This amortizes environment interaction and improves the sample efficiency of best-response computation. Because JBR converts BR computation into an offline RL problem, we propose three remedies for distribution-shift bias: (i) Conservative JBR with safe policy improvement, (ii) Exploration-Augmented JBR that perturbs data collection and admits theoretical guarantees, and (iii) Hybrid BR that interleaves JBR with periodic independent BR updates. Across benchmark multi-agent environments, Exploration-Augmented JBR achieves the best accuracy-efficiency trade-off, while Hybrid BR attains near-PSRO performance at a fraction of the sample cost. Overall, JBR makes PSRO substantially more practical for large-scale strategic learning while preserving equilibrium robustness.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑