Prompt Decision Transformer Based Policy Generalization for Base Station Energy Saving

In the context of 5G networks, the proliferation of base stations (BSs), expansion of signal coverage, and the demand for high data rates and low latency pose significant challenges in reducing energy consumption and enhancing overall energy efficiency. While traditional reinforcement learning (RL) holds promise for managing BSs energy consumption, it necessitates extensive environmental interaction and training time. Moreover, changes in state and action space will result in low sample efficiency and limited generalization ability. In order to address these issues, we propose an innovative approach, namely Prompt Decision Transformer (PDT), which leverages historical empirical data for learning. Additionally, prompts are incorporated into the model to assist the trajectory return, allowing Decision Transformer (DT) to achieve the desired high return, with the state and action of each BS being decoupled to facilitate flexible task switching. Numerical simulation results demonstrate that PDT exhibits robust learning capability and performs well in energy saving of BSs, and can adjust to the new task with a variable number of BSs without retraining, showcasing its strong generalization ability.