A Comparative Study of Deep Reinforcement Learning-Based Downlink Power Control

In rapidly evolving urban environments, cellular networks oftentimes experience inefficiencies in energy utilization while maintaining connections to user equipment (UEs), predominantly mobile phones. Conventional power control algorithms have been employed to mitigate energy waste; however, these algorithms lack in developing network systems, and some rely on requests from UEs that cause latency during connection establishment. In the context of contemporary advancements, artificial intelligence (AI) presents promising solutions for different application fields, particularly through the application of reinforcement learning algorithms for control operations. This study builds upon the research on downlink power control for voice over LTE in [1]. In that paper, a closed-loop framework leveraging conventional reinforcement learning for the power control of antennas which operates independently of power commands from UEs was proposed. This paper extends this previous work by leveraging advanced reinforcement learning algorithms that are Advantage Actor-Critic (A2C), Deep-Q Network (DQN) and Proximal Policy Optimization (PPO). Among the algorithms, Deep-Q Network yields improved outcomes in downlink power control efficiency by considering voice retainability metric as compared to other deep reinforcement learning algorithms as well as original research outcomes. The results obtained also show that communication systems may reap benefits of AI by increasing energy efficiency, network performance, and sustainability.