Prompt Multi-Agent Decision Transformer for Energy-Saving Heterogeneous Networks

Abstract

With the evolution toward sixth-generation (6G) wireless networks, energy efficiency has become a foundational design principle. This paper proposes Prompt Multi-Agent Decision Transformer (PMADT) for energy-efficient sleep mode control in heterogeneous cellular networks. Unlike existing multi-agent decision transformers that rely on implicit coordination, PMADT introduces an interference-aware prompt mechanism that explicitly encodes bidirectional interference relationships between small base stations (SBSs). A key advantage is that these prompts can be computed retroactively from existing trajectory data without requiring new data collection. The prompts are aggregated into fixed-dimensional vectors through descending-order sorting and zero-padding, decoupling model architecture from network scale and enabling crossscale generalization. The proposed offline-to-online framework first pre-trains the model on expert data collected by multiagent proximal policy optimization (MAPPO), and then finetunes the policy through online interactions. Simulation results demonstrate that PMADT achieves superior energy efficiency and quality of service (QoS) compared to baseline methods while generalizing to unseen network configurations.