STAIRS-Former: Spatio-Temporal Attention with Interleaved Recursive Structure Transformer for Offline Multi-task Multi-agent Reinforcement Learning

Jiwon Jeon; Myungsik Cho; Youngchul Sung

cs.AIMar 12, 2026

STAIRS-Former: Spatio-Temporal Attention with Interleaved Recursive Structure Transformer for Offline Multi-task Multi-agent Reinforcement Learning

Authors: Jiwon Jeon, Myungsik Cho, Youngchul Sung

Organizations: School of Electrical Engineering Korea Advanced Institute of Science and Technology (KAIST) Daejeon 34141, Republic of Korea

Abstract

Offline multi-agent reinforcement learning (MARL) with multi-task datasets is challenging due to varying numbers of agents across tasks and the need to generalize to unseen scenarios. Prior works employ transformers with observation tokenization and hierarchical skill learning to address these issues. However, they underutilize the transformer attention mechanism for inter-agent coordination and rely on a single history token, which limits their ability to capture long-horizon temporal dependencies in partially observable MARL settings. In this paper, we propose STAIRS-Former, a transformer architecture augmented with spatial and temporal hierarchies that enables effective attention over critical tokens while capturing long interaction histories. We further introduce token dropout to enhance robustness and generalization across varying agent populations. Extensive experiments on diverse multi-agent benchmarks, including SMAC, SMAC-v2, MPE, and MaMuJoCo, with multi-task datasets demonstrate that STAIRS-Former consistently outperforms prior methods and achieves new state-of-the-art performance.

Figures & tables

Figure 1: Overall Proposed Q Structure

Panel of Figure 2 — Figure 2: Attention map on both seen and unseen task with basic transformer in HiSSD

Figure 3: Overview of STAIRS-Former architecture.

Panel of Figure 4 — Figure 4: Attention map on both seen and unseen task with basic transformer in Ours.

Tasks	Expert				Medium
Tasks	UPDeT-m	ODIS	HiSSD	STAIRS (Ours)	UPDeT-m	ODIS	HiSSD	STAIRS (Ours)
Source Tasks
3m	$66.9\pm 22.7$	$98.1\pm 1.7$	$\bm{99.4\pm 1.4}$	$\bm{99.4\pm 1.4}$	$33.8\pm 21.4$	$59.4\pm 12.3$	$65.0\pm 11.1$	$\bm{84.4\pm 4.4}$
5m6m	$6.9\pm 7.1$	$43.1\pm 28.2$	$\bm{72.5\pm 10.7}$	$70.6\pm 10.5$	$1.9\pm 2.8$	$22.5\pm 10.0$	$35.6\pm 9.5$	$\bm{50.0\pm 12.5}$
9m10m	$24.4\pm 24.4$	$55.6\pm 28.2$	$\bm{99.4\pm 1.4}$	$\bm{99.4\pm 1.4}$	$31.3\pm 31.2$	$57.5\pm 21.5$	$68.1\pm 13.5$	$\bm{86.9\pm 7.5}$
Unseen Tasks

Table 1: Comparison of average and per-task performances on the Marine-Hard task set across four dataset qualities. We report mean

\pm

standard deviation, with the best shown in bold .

Tasks	Expert				Medium
Tasks	UPDeT-m	ODIS	HiSSD	STAIRS (Ours)	UPDeT-m	ODIS	HiSSD	STAIRS (Ours)
Source Tasks
2s3z	$40.6\pm 33.9$	$56.9\pm 29.3$	$90.6\pm 7.7$	$\bm{95.6\pm 5.2}$	$27.5\pm 15.5$	$44.4\pm 16.1$	$46.9\pm 15.5$	$\bm{56.9\pm 10.5}$
2s4z	$16.3\pm 14.6$	$52.5\pm 18.8$	$\bm{80.0\pm 10.5}$	$77.5\pm 11.6$	$21.3\pm 21.8$	$21.3\pm 11.3$	$15.0\pm 8.1$	$\bm{60.0\pm 16.1}$
3s5z	$23.8\pm 27.2$	$65.6\pm 37.8$	$\bm{90.6\pm 3.8}$	$87.5\pm 10.6$	$13.1\pm 9.5$	$15.6\pm 8.0$	$28.1\pm 20.1$	$\bm{52.5\pm 3.4}$
Unseen Tasks

Table 2: Comparison of average and per-task performances on the Stalker-Zealot task set across four dataset qualities. We report mean

\pm

standard deviation, with the best shown in bold .

	Tasks	UPDeT-m	ODIS	HiSSD	STAIRS (Ours)
Seen	Marine-Hard	21.2	47.9	64.6	79.0
	Marine-Easy	44.3	59.3	83.9	91.2
	Stalker-Zealot	20.3	34.8	45.9	63.4
	Mean	28.6	47.3	64.8	77.9
Unseen	Marine-Hard	21.1	31.8	52.7	57.0
	Marine-Easy	29.9	42.5	79.8	86.7

Table 3: Results on seen and unseen tasks, averaged over dataset quality. Best in bold , second-best underlined .

	Tasks	UPDeT-m	ODIS	HiSSD	STAIRS (Ours)
Seen	Terran	10.9	15.2	24.9	31.3
	Protoss	9.9	12.7	29.5	32.5
	Zerg	6.6	10.4	21.0	29.2
	Mean	9.1	12.7	25.1	31.0
Unseen	Terran	7.0	13.2	25.1	32.3
	Protoss	8.1	11.6	28.5	32.8

Table 4: Results on SMAC-v2 benchmark. Best in bold , second-best underlined .

Figure 5: Temporal attention map in a SMAC 3m scenario. The attention maps from STAIRS-Former (top) and HiSSD (bottom) illustrate how attention shifts over time. Lighter-colored regions indicate eliminated agents. A detailed explanation of these heatmaps is provided in Appendix G.1

	Tasks	STAIRS	w/o Temporal	w/o Spatial	w/o Dropout	w/o ST	w/o STD
Seen	Marine-Hard	79.0	77.5	71.0	75.7	69.9	66.0
	Marine-Easy	91.2	88.1	87.2	89.6	86.9	87.9
	Stalker-Zealot	63.4	63.0	59.0	62.6	50.3	55.1
	Mean	77.9	76.2	72.4	76.0	69.0	69.6
Unseen	Marine-Hard	57.0	57.4	54.7	56.0	54.1	40.1
	Marine-Easy	86.7	78.5	79.0	83.0	78.0	79.7

Table 5: Ablation results on Seen and Unseen tasks. “ST” = Spatial & Temporal, “STD” = ST + Dropout. The best performance is shown in bold , and the second-best performance is underlined .

Panel of Figure 6 — Figure 6: Dormant neuron ratios on marine-hard tasks: STAIRS vs. ablations.

Appendix figures & tables33 assets

Supplementary material from the paper’s appendix.

Appendix

Task	Quality	# Trajectories	Average return	Average win rate
3m	expert	2000	19.89	0.99
	medium	2000	13.99	0.54
	medium-expert	4000	16.94	0.77
	medium-replay	3630	N/A	N/A
5m	expert	2000	19.94	0.99
	medium	2000	17.33	0.74

Appendix

Table 6: Properties of offline datasets with different qualities.

Task type	Task	Ally units	Enemy units	Properties
Source	3m	3 Marines	3 Marines	homogeneous & symmetric
	5m	5 Marines	5 Marines	homogeneous & symmetric
	10m	10 Marines	10 Marines	homogeneous & symmetric
Unseen	4m	4 Marines	4 Marines	homogeneous & symmetric
	6m	6 Marines	6 Marines	homogeneous & symmetric
	7m	7 Marines	7 Marines	homogeneous & symmetric

Appendix

Table 7: Descriptions of marine-easy tasks

Task type	Task	Ally units	Enemy units	Properties
Source	3m	3 Marines	3 Marines	homogeneous & symmetric
	5m_vs_6m	5 Marines	6 Marines	homogeneous & asymmetric
	9m_vs_10m	9 Marines	10 Marines	homogeneous & asymmetric
Unseen	4m	4 Marines	4 Marines	homogeneous & symmetric
	5m	5 Marines	5 Marines	homogeneous & symmetric
	10m	10 Marines	10 Marines	homogeneous & symmetric

Appendix

Table 8: Descriptions of marine-hard tasks

Task type	Task	Ally units	Enemy units	Properties
Source	2s3z	2 Stalkers, 3 Zealots	2 Stalkers, 3 Zealots	heterogeneous & symmetric
	2s4z	2 Stalkers, 4 Zealots	2 Stalkers, 4 Zealots	heterogeneous & symmetric
	3s5z	3 Stalkers, 5 Zealots	3 Stalkers, 5 Zealots	heterogeneous & symmetric
Unseen	1s3z	1 Stalker, 3 Zealots	1 Stalker, 3 Zealots	heterogeneous & symmetric
	1s4z	1 Stalker, 4 Zealots	1 Stalker, 4 Zealots	heterogeneous & symmetric
	1s5z	1 Stalker, 5 Zealots	1 Stalker, 5 Zealots	heterogeneous & symmetric

Appendix

Table 9: Descriptions of stalker-zealot tasks

Tasks	Expert				Medium
Tasks	UPDeT-m	ODIS	HiSSD	STAIRS (Ours)	UPDeT-m	ODIS	HiSSD	STAIRS (Ours)
Source Tasks
3m	$58.8\pm 20.1$	$84.4\pm 18.9$	$\bm{100.0\pm 0.0}$	$99.4\pm 1.4$	$55.6\pm 28.8$	$60.0\pm 6.8$	$67.5\pm 10.5$	$\bm{85.6\pm 6.5}$
5m	$48.8\pm 35.7$	$79.4\pm 29.0$	$\bm{99.4\pm 1.4}$	$\bm{99.4\pm 1.4}$	$71.9\pm 7.3$	$79.4\pm 10.3$	$80.6\pm 6.4$	$\bm{85.0\pm 9.2}$
10m	$46.3\pm 36.4$	$67.5\pm 41.8$	$\bm{99.4\pm 1.4}$	$\bm{99.4\pm 1.4}$	$48.8\pm 27.9$	$77.5\pm 13.7$	$66.3\pm 19.6$	$\bm{94.4\pm 2.6}$
Unseen Tasks

Appendix

Table 10: Comparison of average and per-task performances on the Marine-Easy task set across four dataset qualities. We report mean

\pm

standard deviation, with the best shown in bold .

Task	Quality	# Trajectories	Average return	Average win rate
CN-2	expert	2000	1.0000	1.0000
CN-2	medium	2000	0.6152	0.6152
CN-4	expert	2000	0.7173	0.7173
CN-4	medium	2000	0.4273	0.4273

Appendix

Table 11: Properties of offline datasets with different qualities.

Tasks	Expert
Tasks	UPDeT-m	ODIS	HiSSD	STAIRS (Ours)
Source Tasks
CN-2	$68.8\pm 19.4$	$78.8\pm 44.1$	$\bm{100.0\pm 0.0}$	$\bm{100.0\pm 0.0}$
CN-4	$13.8\pm 12.6$	$21.9\pm 15.1$	$24.4\pm 1.4$	$\bm{30.0\pm 12.6}$
Unseen Tasks
CN-3	$34.4\pm 19.1$	$48.8\pm 27.8$	$\bm{65.0\pm 10.2}$	$64.4\pm 5.7$

Appendix

Table 12: Comparison of average and per-task performances on the Cooperative navigation task set across two dataset qualities. We report mean

\pm

standard deviation, with the best shown in bold .

Figure 7: Structue of attention map

Figure 8: Another temporal evolution in a SMAC 3m episode: STAIRS-Former (Above) vs. HiSSD (Below)

Figure 9: Attention maps and trajectories on other tasks: STAIRS-Former (Above) vs. HiSSD (Below)

Figure 10: Average attention map on marine-hard task with HiSSD

Figure 11: Average attention map on stalker-zealot task with HiSSD

Figure 12: Average attention map on marine-easy task with HiSSD

Figure 13: Average attention map on marine-hard task with STAIRS

Figure 14: Average attention map on stalker-zealot task with STAIRS

Figure 15: Average attention map on marine-easy task with STAIRS

Hyper-parameter	Value
hidden layer dimension	64
attention dimension	64
$\lambda$	1.0
optimizer	Adam
learning rate	0.0005
Number of layers $M$	$2$

Appendix

Table 13: Hyper-parameters of STAIRS-Former

Algorithms	UpDeT-m	ODIS	HiSSD	STAIRS
# Parameters	79,095	138,573	679,335	220,023
GPU Memory Usage (MiB)	7,046	7,020	17,492	14,370

Appendix

Table 14: The parameter counts and GPU memory footprint for UpDeT-m, ODIS, HiSSD, and our method (STAIRS) on the Marine-Hard-Medium task

	Temporal Interval $\mathbf{T_{H}}$
	$3$		$4$		$5$
	Token Dropout Rate $\mathbf{p_{\text{drop}}}$
	0.05	0.1	0.05	0.1	0	0.05	0.1
Average Performance	65.9	65.3	64.5	66.2	63.6	66.2	65.6

Appendix

Table 15: Comparison of average performances over all task set and dataset qualities. Best performance are shown in bold .

Task	Quality	# Trajectories	Average return	Average win rate
Terran 3_vs_3	medium	2000	11.6	0.44
Terran 3_vs_3	medium-replay	2000	N/A	N/A
Terran 5_vs_5	medium	2000	13.09	0.4
Terran 5_vs_5	medium-replay	2000	N/A	N/A
Terran 10_vs_10	medium	2000	12.58	0.42
Terran 10_vs_10	medium-replay	2000	N/A	N/A

Appendix

Table 16: Properties of offline datasets with different qualities.

Race	Unit Types (weights)	Start Pos. Dist.
Terran	marine (0.45), marauder (0.45), medivac (0.10)	surrounded_and_reflect ( $p{=}0.5$ )
Protoss	stalker (0.45), zealot (0.45), colossus (0.10)	surrounded_and_reflect ( $p{=}0.5$ )
Zerg	zergling (0.45), baneling (0.10), hydralisk (0.45)	surrounded_and_reflect ( $p{=}0.5$ )

Appendix

Table 17: Unit generation and start-position configuration for each SMAC-v2 race.

Tasks	Medium				Medium-replay
Tasks	UPDeT-m	ODIS	HiSSD	STAIRS (Ours)	UPDeT-m	ODIS	HiSSD	STAIRS (Ours)
Terran Source Tasks
Terran 3	$15.0\pm 2.6$	$18.1\pm 9.2$	$31.3\pm 11.0$	$\bm{37.5\pm 5.8}$	$10.6\pm 5.2$	$19.4\pm 12.2$	$\bm{31.3\pm 8.6}$	$28.1\pm 15.9$
Terran 5	$16.3\pm 6.0$	$16.3\pm 6.8$	$18.8\pm 3.1$	$\bm{26.9\pm 5.7}$	$8.8\pm 4.1$	$11.9\pm 10.0$	$\bm{23.1\pm 11.0}$	$31.3\pm 6.3$
Terran 10	$10.6\pm 9.3$	$15.0\pm 11.1$	$22.5\pm 11.1$	$\bm{36.9\pm 9.7}$	$3.8\pm 2.6$	$10.0\pm 10.2$	$22.5\pm 8.4$	$\bm{26.9\pm 17.5}$
Terran Unseen Tasks

Appendix

Table 18: Comparison of average and per-task performances on the SMAC-V2 task set. We report mean

\pm

standard deviation, with the best result shown in bold . For brevity, we abbreviate task names such as 3_vs_3 to Terran 3 , Protoss 3 , and so on.

Task type	Task	Number of Agents	Observation Space	Action Space
Source	(3,3)	2	[(8,), (8,)]	[(3,), (3,)]
	(2,2,2)	3	[(6,), (10,), (8,)]	[(2,), (2,), (2,)]
	(1,1,1,1,1,1)	6	[(4,), (6,), (6,), (4,), (6,), (6,)]	[(1,)] $\times$ 6
Unseen	(6)	1	[(12,)]	[(6,)]
	(2,4)	2	[(6,), (10,)]	[(2,), (4,)]
	(1,2,3)	3	[(4,), (8,), (8,)]	[(1,), (2,), (3,)]

Appendix

Table 19: Descriptions of ’HalfCheetah’

Task	Quality	# Trajectories	Average return
(3,3)	medium	100	5043.32
(2,2,2)	medium	100	5074.4
(1,1,1,1,1,1)	medium	100	4076.55

Appendix

Table 20: Properties of offline datasets on ’HalfCheetah’

Tasks	HalfCheetah
Tasks	UPDeT-BC	ODIS	STAIRS (Ours)
	Source Tasks
(3,3)	$148.2\pm 307.9$	$970.4\pm 416.9$	$\bm{1459.0\pm 400.3}$
(2,2,2)	$-66.4\pm 124.1$	$537.8\pm 318.8$	$\bm{1410.6\pm 537.6}$
(1,1,1,1,1,1)	$262.6\pm 200.7$	$727.6\pm 662.0$	$\bm{1006.0\pm 420.4}$
	Unseen Tasks

Appendix

Table 21: Comparison of average and per-task performances on the HalfCheetah task set in MAMuJoCo. We report mean

\pm

standard deviation, with the best result shown in bold .

Figure 17: Average attention maps of models trained on the Marine-Hard-Medium task, evaluated on the seen (3m) and unseen (4m) tasks across the entire trajectory for different values of $\tau$ .

Figure 18: Attention maps of models trained on the Marine-Hard-Medium task, evaluated on the 3m task and sampled every 2–3 timesteps for different values of $\tau$ .

Tasks	Expert				Medium
Tasks	UPDeT-m	ODIS	HiSSD	STAIRS (Ours)	UPDeT-m	ODIS	HiSSD	STAIRS (Ours)
Source Tasks
3m	$78.8\pm 8.4$	$94.4\pm 10.9$	$\mathbf{99.4\pm 1.4}$	$\mathbf{99.4\pm 1.4}$	$41.9\pm 29.0$	$61.9\pm 18.5$	$62.5\pm 10.4$	$\mathbf{84.4\pm 4.4}$
5m6m	$5.0\pm 4.7$	$38.1\pm 13.1$	$\mathbf{72.5\pm 6.8}$	$70.6\pm 10.5$	$5.0\pm 11.2$	$28.1\pm 9.1$	$33.8\pm 13.3$	$\mathbf{50.0\pm 12.5}$
9m10m	$18.1\pm 19.3$	$71.9\pm 15.9$	$97.5\pm 1.4$	$\mathbf{99.4\pm 1.4}$	$15.0\pm 16.4$	$51.9\pm 25.1$	$68.8\pm 13.4$	$\mathbf{86.9\pm 7.5}$
Unseen Tasks

Appendix

Table 23: Comparison of average and per-task performances on the Marine-hard task set across four dataset qualitieswith all transformer backbones using depth 2. We report mean

\pm

standard deviation, with the best shown in bold .

Tasks	Expert				Medium
Tasks	UPDeT-m	ODIS	HiSSD	STAIRS (Ours)	UPDeT-m	ODIS	HiSSD	STAIRS (Ours)
Source Tasks
2s3z	$33.8\pm 41.4$	$70.0\pm 38.6$	$92.5\pm 5.7$	$\bm{95.6\pm 5.2}$	$25.0\pm 12.1$	$43.1\pm 19.2$	$39.4\pm 12.4$	$\bm{56.9\pm 10.5}$
2s4z	$13.8\pm 14.3$	$58.8\pm 37.3$	$65.0\pm 5.1$	$\bm{77.5\pm 11.6}$	$25.0\pm 20.6$	$7.5\pm 3.6$	$9.4\pm 4.9$	$\bm{60.0\pm 16.1}$
3s5z	$28.1\pm 24.5$	$66.9\pm 33.0$	$\bm{88.8\pm 5.7}$	$87.5\pm 10.6$	$20.6\pm 12.2$	$24.4\pm 9.7$	$26.8\pm 12.0$	$\bm{52.5\pm 3.4}$
Unseen Tasks

Appendix

Table 24: Comparison of average and per-task performances on the Stalker-Zealot task set across four dataset qualitieswith all transformer backbones using depth 2. We report mean

\pm

standard deviation, with the best shown in bold .

Tasks	Expert				Medium
Tasks	UPDeT-m	ODIS	HiSSD	STAIRS (Ours)	UPDeT-m	ODIS	HiSSD	STAIRS (Ours)
Source Tasks
3m	$71.9\pm 21.1$	$96.3\pm 3.4$	$\bm{100.0\pm 0.0}$	$99.4\pm 1.4$	$63.1\pm 20.4$	$46.9\pm 12.1$	$70.6\pm 5.7$	$\bm{85.6\pm 6.5}$
5m	$55.0\pm 20.4$	$97.5\pm 3.4$	$\bm{100.0\pm 0.0}$	$99.4\pm 1.4$	$73.1\pm 6.8$	$78.1\pm 3.8$	$78.8\pm 1.4$	$\bm{85.0\pm 9.2}$
10m	$48.1\pm 15.7$	$96.3\pm 4.1$	$\bm{100.0\pm 0.0}$	$99.4\pm 1.4$	$56.9\pm 12.8$	$59.4\pm 17.7$	$75.6\pm 9.7$	$\bm{94.4\pm 2.6}$
Unseen Tasks

Appendix

Table 25: Comparison of average and per-task performances on the Marine-easy task set across four dataset qualitieswith all transformer backbones using depth 2. We report mean

\pm

standard deviation, with the best shown in bold .

Task / Dataset	UpDeT-m	UpDeT-m + GRU	UpDeT-bc	UpDeT-bc + GRU	Ours
Marine-Hard	21.1	20.7	46.6	49.8	62.5
Stalker–Zealot	15.3	16.7	43.3	42.9	51.7
Marine-Easy	34.2	31.6	82.1	85.7	88.1

Appendix

Table 26: Average performance comparison of GRU addition.

Figure 19: Average attention maps of models trained on the Marine-Hard-Medium task, evaluated on the seen (3m) and unseen (4m) tasks. UpDeT-m (upper) vs UpDeT-m with an added GRU history token (lower).

Offline MARL

Offline multi-agent reinforcement learning (MARL) learns coordinated policies from fixed interaction datasets without further environment exploration.

See all 30 papers

Bridging MARL to SARL: An Order-Independent Multi-Agent Transformer via Latent Consensus
Apr 15, 2026Zijian Zhao, Jing Gao, Sen LiCooperative MARL Multi-Agent Reinforcement Learning
1The Hong Kong University of Science and Technology · 2The Hong Kong Polytechnic University · 3The Hong Kong University of Science and Technology (Guangzhou)
MARL-GPT: Foundation Model for Multi-Agent Reinforcement Learning
Apr 7, 2026Maria Nesterova, Mikhail Kolosov, Anton Andreychuk +6Multi-Agent Reinforcement Learning Transformer-Based RL
AXXX & MIRAI, Moscow, Russia · MIRAI, Moscow, Russia · AXXX, Moscow, Russia +2
Transformer-Based Multi-Agent Reinforcement Learning for Networked Systems with Long-Range Interactions
Nov 17, 2025Vidur Sinha, Muhammed Ustaomeroglu, Guannan QuCounterfactual Credit Assignment Graph Transformer
Department of Electrical and Computer Engineering, Carnegie Mellon University, Pittsburgh, USA

Apr 15, 2026J/K move · Enter open · S save

Bridging MARL to SARL: An Order-Independent Multi-Agent Transformer via Latent Consensus

Zijian Zhao, Jing Gao, Sen Li

1The Hong Kong University of Science and Technology · 2The Hong Kong Polytechnic University · 3The Hong Kong University of Science and Technology (Guangzhou)

Cooperative MARL Multi-Agent Reinforcement Learning Multi-Agent Coordination Transformer-Based RL Multi-Agent Consensus

Open paper AskPDF arXiv

Cooperative multi-agent reinforcement learning (MARL) is widely used to address large joint observation and action spaces by decomposing a centralized control problem into multiple interacting agents. However, such decomposition often introduces additional challenges, including non-stationarity, unstable training, weak coordination, and limited theoretical guarantees. In this paper, we propose the Consensus Multi-Agent Transformer (CMAT), a centralized framework that bridges cooperative MARL to a hierarchical single-agent reinforcement learning (SARL) formulation. CMAT treats all agents as a unified entity and employs a Transformer encoder to process the large joint observation space. To handle the extensive joint action space, we introduce a hierarchical decision-making mechanism in which a Transformer decoder autoregressively generates a high-level consensus vector, simulating the process by which agents reach agreement on their strategies in latent space. Conditioned on this consensus, all agents generate their actions simultaneously, enabling order-independent joint decision making and avoiding the sensitivity to action-generation order in conventional Multi-Agent Transformers (MAT). This factorization allows the joint policy to be optimized using single-agent PPO while preserving expressive coordination through the latent consensus. To evaluate the proposed method, we conduct experiments on benchmark tasks from StarCraft II, Multi-Agent MuJoCo, and Google Research Football. The results show that CMAT achieves superior performance over recent centralized solutions, sequential MARL methods, and conventional MARL baselines. The code for this paper is available at:https://github.com/RS2002/CMAT .

Abstract

Offline MARL

Explore similar work

Bridging MARL to SARL: An Order-Independent Multi-Agent Transformer via Latent Consensus

MARL-GPT: Foundation Model for Multi-Agent Reinforcement Learning

Transformer-Based Multi-Agent Reinforcement Learning for Networked Systems with Long-Range Interactions