Transformer‐Based Fusion for Joint Routing and Resource Allocation in the Internet of Things

ABSTRACT High‐volume data transmission in Internet of Things (IoT) networks demands the establishment of high‐quality multi‐hop communication paths at the network layer. To address this challenge, this paper proposes a joint routing and resource allocation framework that simultaneously optimizes relay node selection and transmit power allocation under decentralized manner. The proposed approach leverages deep reinforcement learning (DRL) to enable hop‐by‐hop decision‐making with only local observations. To mitigate the partial observability stemmed from limited perception range of node, a transformer‐based architecture is integrated into the DRL agent to enable the fusion of observations and actions collected along the established path. Specifically, the encoder module aggregates historical and frontier observations along the forwarding path, while the decoder module models temporal dependencies among historical actions. The fused representation is utilized to generate optimal decisions for next‐hop node selection and power allocation. After each action execution, a one‐hot encoded placeholder of the chosen action is fed back into the decoder module for subsequent decisions. Finally, numerical simulations demonstrate that the proposed transformer‐enhanced DRL framework significantly outperforms state‐of‐the‐art baselines in terms of end‐to‐end path quality, and robustness under decentralized IoT network configuration.