MASTARS: Multi-Agent Sequential Trajectory Augmentation with Return-Conditioned Subgoals
MASTARS generates training trajectories one agent at a time to preserve both variety and coordination.
Offline multi-agent reinforcement learning is constrained by the quality of its fixed training dataset. The authors describe a trade-off: generating agents’ trajectories independently adds diversity but weakens coordination, while generating them jointly tends to reproduce the training distribution. Their diffusion-based framework uses inpainting to keep sequentially generated trajectories consistent across agents. They report significant performance gains across multiple offline learning methods on MPE, SMAC and SMAC-v2 benchmarks; the abstract gives no numerical results.
ArXiv · AI/CL/LG's note
Offline multi-agent reinforcement learning is constrained by the quality of its fixed training dataset. The authors describe a trade-off: generating agents’ trajectories independently adds diversity but weakens coordination, while generating them jointly tends to reproduce the training distribution. Their diffusion-based framework uses inpainting to keep sequentially generated trajectories consistent across agents. They report significant performance gains across multiple offline learning methods on MPE, SMAC and SMAC-v2 benchmarks; the abstract gives no numerical results.
ArXiv · AI/CL/LG's note
score 4