Attention-based Hierarchical Variational Information Bottleneck for Robust Multi-Agent Communication under Variable Bandwidth

Lukas Koch Vindbjerg; Qi Zhang; Yury Brodskiy; Lukas Esterle

cs.LGSep 28, 2026

Attention-based Hierarchical Variational Information Bottleneck for Robust Multi-Agent Communication under Variable Bandwidth

Authors: Lukas Koch Vindbjerg, Qi Zhang, Yury Brodskiy, Lukas Esterle

Organizations: Department of Electrical and Computer Engineering, Aarhus University, Aarhus, Denmark · EIVA a/s, Denmark · DIGIT, Aarhus University, Aarhus, Denmark

Abstract

Learning-based multi-agent communication under limited bandwidth does not only require deciding what to communicate, but also structuring messages so that partial transmissions remain useful. We study this problem under prefix truncation, where only the first part of each message is received. To address it, we propose \textbf{AH-VIB}, an attention-based autoregressive variational communication model that combines a variational information bottleneck (VIB) with sequential message generation and a hierarchical robustness loss. We evaluate AH-VIB on a custom cooperative object-inspection and occupancy-mapping task, where agents equipped with a limited field-of-view sensor coordinate to scan inspection objects in an occupancy-grid world, under variable and fixed bandwidth conditions, and compare it against MADDPG, CommNet, a flat VIB baseline, and an autoregressive MLP ablation. AH-VIB achieves competitive mean return while improving performance reliability under the most constrained bandwidth conditions. These results indicate that AH-VIB improves the reliability and graceful degradation of learned communication under bandwidth constraints.

Figures & tables

Fig. 1: Cooperative object inspection under partial observability. Agents with limited fields of view progressively reveal free space and inspection-object surfaces while coordinating their coverage.

Hyperparameter	Value	Hyperparameter	Value	Hyperparameter	Value
Environment & Training		Optimisation		Exploration Noise
Total env. steps	$5\times 10^{5}$	Actor LR	$3\times 10^{-4}$	Initial noise $\sigma_{0}$	$0.20$
Episode length $T$	$200$	Critic LR	$3\times 10^{-5}$	Decay factor	$0.80$
Replay buffer size	$2\times 10^{5}$	Discount $\gamma$	$0.95$	Decay over (steps)	$1\times 10^{5}$
Batch size	$256$	Target network $\tau$	$0.005$	Min. noise $\sigma_{\min}$	$0.05$
Random exploration steps	$4{,}000$	Actor hidden layers	$[512,\;256,\;256,\;128]$

TABLE I: Training hyperparameters used in the experiments, shared across all algorithms unless noted.

Algorithm	VIB	AR	Hier.	Attn.
MADDPG	$\times$	$\times$	$\times$	$\times$
CommNet	$\times$	$\times$	$\times$	$\times$
VIB	✓	$\times$	$\times$	$\times$
HVIB	✓	✓	✓	$\times$
AH-VIB	✓	✓	✓	✓

TABLE II: Baseline design comparison.

	$\rho$ =0.10	$\rho$ =0.25	$\rho$ =0.75	$\rho$ =1.0
AH-VIB	-0.2 $\pm$ 0.5	-0.2 $\pm$ 0.5	-0.4 $\pm$ 1.9	-0.4 $\pm$ 2.2
VIB	-7.7 $\pm$ 24.3	-6.2 $\pm$ 21.6	-6.1 $\pm$ 21.6	-7.3 $\pm$ 23.0
MADDPG	-5.4 $\pm$ 16.7	-5.3 $\pm$ 16.9	-13.7 $\pm$ 29.2	-1.1 $\pm$ 8.4
COMMNET	-1.8 $\pm$ 10.1	-2.1 $\pm$ 12.2	-0.1 $\pm$ 0.3	-0.1 $\pm$ 0.3

TABLE III: Episode return at selected bandwidth fractions (mean

\pm

standard deviation). Bold indicates the best mean and lowest standard deviation in each column.

Explore similar work

Decoupling Communication from Policy: Robust MARL under Bandwidth Constraints
May 20, 2026Alexi Canesse, Benoît Goupil, Jesse Read +1Multi-Agent Reinforcement Learning Bandwidth Extension
École polytechnique (LIX), CNRS, Institut Polytechnique de Paris, Palaiseau, France
HiComm: Hierarchical Communication for Multi-agent Reinforcement Learning
Jun 28, 2026Runze Zhao, Dongruo Zhou, Sumit Kumar Jha +2Multi-Agent Reinforcement Learning Vectors
Luddy School of Informatics, Computing, and Engineering Indiana University Bloomington Bloomington, IN 47408 · Department of Computer & Information Science & Engineering University of Florida Gainesville, FL 32611 · Department of Electrical Engineering and Computer Science United States Military Academy West Point, NY 10996 +1
Robust and Efficient Communication for Multi-Agent Learning
Sep 14, 2026Rafael Pina, Varuna De Silva, Corentin ArtaudMulti-Agent Reinforcement Learning Artificial Intelligence Agents
Institute for Digital Technologies · Loughborough University London, United Kingdom

May 20, 2026, cs.MAJ/K move · Enter open · S save

Decoupling Communication from Policy: Robust MARL under Bandwidth Constraints

Alexi Canesse, Benoît Goupil, Jesse Read, Sonia Vanier

École polytechnique (LIX), CNRS, Institut Polytechnique de Paris, Palaiseau, France

Multi-Agent Reinforcement Learning Bandwidth Extension Decoupling Capacity

Open paper AskPDF arXiv

Communication enables coordination in multi-agent reinforcement learning (MARL), but many real-world applications, e.g., search-and-rescue with drone swarms, operate under severe bandwidth constraints. Many communication architectures still expose a coupled bottleneck in which a shared latent representation is used for both policy execution and inter-agent communication. Consequently, reducing message size directly limits the policy's latent space, often leading to significant performance degradation. We address this with two contributions. First, we introduce

β

, a normalised per-agent bandwidth budget that unifies sparsity, rounds, and message dimension into a single comparable constraint. Second, we provide SLIM, a minimal architecture that decouples the communication pathway from the policy's latent representation, allowing us to isolate the effect of bandwidth from the effect of policy capacity while benefiting from in-step communication. We evaluate our method on several partially-observable MARL benchmarks, where communication is essential. Our approach achieves state-of-the-art performance and exhibits scalability and robustness under limited communication, with only marginal degradation as bandwidth is reduced.

Abstract

Explore similar work

Decoupling Communication from Policy: Robust MARL under Bandwidth Constraints

HiComm: Hierarchical Communication for Multi-agent Reinforcement Learning

Robust and Efficient Communication for Multi-Agent Learning