Design and Evaluation of Intelligent Multi-Agent Systems for Complex Networks

Abstract

Complex networks — encompassing communication infrastructure, smart energy grids, logistics systems, and social interaction platforms — present
inherent challenges of scale, heterogeneity, and dynamic topology that render centralised control intractable. Intelligent multi-agent systems (MAS)
offer a decentralised paradigm in which autonomous agents cooperate and coordinate to achieve global objectives through local interactions, without
requiring complete global state knowledge. This study presents NetMAS, a scalable intelligent multi-agent framework incorporating decentralised
partially observable Markov decision processes (Dec-POMDP), graph neural network (GNN)-based communication, and a consensus-driven
coordination protocol for adaptive control of complex networks. NetMAS was evaluated across four complex network domains: 5G heterogeneous
network resource allocation (n = 512 base stations), smart grid demand response (n = 1,024 prosumer nodes), urban logistics route optimisation (n =
180 vehicles, Amsterdam road network), and epidemic containment simulation (n = 100,000 nodes, scale-free network). NetMAS achieved
throughput improvements of 31.7%, cost reductions of 28.4%, delivery efficiency gains of 24.9%, and containment effectiveness improvements of
37.2% over state-of-the-art baselines. GNN-based communication reduced inter-agent message overhead by 58.3% while maintaining coordination
quality. These results establish NetMAS as a scalable and effective architecture for intelligent multi-agent control of large-scale complex networks.