TY  - EJOU
AU  - Zhu, Hua-Yu 
AU  - Mao, Weijie 

TI  - Dimensionality-Adaptive Reinforcement Learning for Energy-Efficient Connectivity-Maintaining Consensus of Multi-Agent Systems with Hybrid Attacks
T2  - Computers, Materials \& Continua

PY  - 
VL  - 
IS  - 
SN  - 1546-2226

AB  - Achieving secure and energy-efficient connectivity-maintaining consensus is a quintessential challenge in multi-agent systems (MAS), given the complex coupling between global connectivity constraints, power limitations, and adversarial attacks. This paper introduces a resilient closed-loop framework that synergizes topology maintenance and consensus control across two time-scales, incorporating real-time fault detection and post-attack autonomous survival protocols. To reconstruct secure and energy-efficient topologies, we develop a dimensionality-adaptive reinforcement learning (RL) scheme, which adaptively deploys pre-trained RL topology reconstruction policies tailored to the spatial dimensions of the current operating environment. To obtain these multi-dimensional policies, we propose an efficient topology reconstruction training mechanism. By sequentially assigning values only to the upper triangular elements of the adjacency matrix, this mechanism exploits network symmetry to compress the action search space and strongly penalizes disconnected configurations. The simulations test our proposed method on energy-efficient resilient consensus problems with 10 agents and cyber-physical attacks in 5 s, and we show that the proposed method matches the state of the art on performance while adding targeted-decapitation survivability and dimensional adaptivity.
KW  - Multi-agent systems; reinforcement learning; consensus; connectivity maintenance; hybrid attacks; energy efficiency

DO  - 10.32604/cmc.2026.086986