Efficient Multi-agent Offline Coordination via Diffusion-based Trajectory Stitching

Authors: Lei Yuan, Yuqi Bian, Lihe Li, Ziqian Zhang, Cong Guan, Yang Yu

ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Empirical results on imbalanced datasets of multiple benchmarks demonstrate that MADi TS significantly improves MARL performance. 5 EXPERIMENT
Researcher Affiliation Collaboration Lei Yuan1,2,3 Yuqi Bian1,2 Lihe Li1,2 Ziqian Zhang1,2 Cong Guan1,2 Yang Yu1,2,3 1 National Key Laboratory for Novel Software Technology, Nanjing University 2 School of Artificial Intelligence, Nanjing University 3 Polixir Technologies
Pseudocode Yes Specifically, we implement our method to the behavior cloning (BC) (Song et al., 2018), OMIGA (Wang et al., 2023) and CFCQL (Shao et al., 2023), and the detailed pseudocode could be found in Appendix F. F PSEUDOCODE
Open Source Code No The paper mentions using 'the Offline MARL framework Off Py MARL codebase (Zhang, 2023) 1and default hyperparameters from Git Hub to implement these algorithms', referring to evaluation algorithms, not explicitly stating the release of the MADi TS implementation itself.
Open Datasets Yes We evaluate our method on two widely-used multi-agent benchmarks that require agent cooperation: the Multi-Agent Particle Environment (MPE) (Lowe et al., 2017), the Star Craft Multi-Agent Challenges (SMAC) (Samvelyan et al., 2019), SMACv2 (Ellis et al., 2023), and MAMu Jo Co (Peng et al., 2021).
Dataset Splits No The paper describes the collection of '40k, 20k, 10k, 2k trajectories in MPE, SMAC, SMACv2, MAMu Jo Co, respectively' and the creation of 'augmented datasets', but does not explicitly detail the training, validation, or test splits for these datasets used in the experiments.
Hardware Specification Yes The experiments were conducted on servers outfitted with Ge Force RTX 2080 Ti.
Software Dependencies No The paper mentions using 'the Offline MARL framework Off Py MARL codebase (Zhang, 2023)' but does not provide specific version numbers for any software dependencies like programming languages or libraries.
Experiment Setup Yes The paper includes 'Table 2: Environment-independent hyperparameters' and 'Table 3: Environment-independent hyperparameters' which list specific values for various parameters such as 'lr 2e-4', 'training steps 1e6', 'batch size for training 32', 'δrecon 0.01', etc.