Efficient Task Offloading in AI-Agent Communication Networks: A Joint Latency and Energy Optimization Approach

Efficient task offloading in dynamic, collaborative multiagent networks is critical for latency-sensitive and energy-constrained applications, yet it presents a significant optimization challenge due to its combinatorial complexity and the non-Euclidean nature of network topologies. While traditional methods struggle with scalability and dynamics, existing learning-based approaches often fail to generalize across diverse network structures. The primary aim of this study is to propose and validate the hierarchical graph transformer (HGT), a novel neural architecture designed to learn powerful and generalizable policies for such complex, mixed-variable optimization problems. Instead of decoupling the problem, the HGT approaches the joint optimization in a holistic, end-to-end manner. By representing the network as a graph, HGT first employs edge-aware graph attention networks (EAGANs) to encode rich and localized physical-layer information into node embeddings. The graph transformer (GT) subsequently processes these embeddings to capture the global context and latent interagent competitive relationships. The resulting context-aware representations are then decoded into a complete action through a multihead architecture that is designed to handle hybrid discrete–continuous action spaces with global constraints. We demonstrate the efficacy of the HGT on the problem of joint latency and energy optimization in collaborative multiagent networks. Compared with strong heuristic, iterative, and recent learning-based baselines, the proposed policy achieves the best weighted utility and latency across the tested task-composition range while remaining highly competitive in energy, and it also shows strong zero-shot transferability to networks of varying scales and topologies. Our work signifies a tangible step towards creating more autonomous and efficient wireless systems, with direct implications for fields like swarm robotics and the Internet of Things.