Wireless Multi-Interface Connectivity with Deep Learning-Enabled User Devices: An Energy Efficiency Perspective

Undoubtedly, future Beyond 5G (B5G) and 6G networks will be empowered by Artificial Intelligence (AI) and Machine Learning (ML) technologies. However, current Deep Learning methods consume tremendous power, making them unsuitable to be deployed at edge Access Points (AP) or user devices. As energy efficiency will become one of the major Key Performance Indicators in 6G, there is an urge for realizing low-power yet highly efficient Deep Learning at user devices, the most severely battery-limited entities. We analyze this issue through the problem of distributed user-to-multiple APs association optimization in B5G networks that integrate diverse wireless interfaces, such as Sub-6GHz and mmWave bands. In the proposed Deep Reinforcement Learning (DRL) framework, each device is equipped by a light-weight Deep Q-Network (DQN), which enables it to self-optimize its APs and interfaces association to fulfill the stringent Quality of Service (QoS) requirements of its various applications. We analyze the energy consumption of the device DQN, and investigate key performance trade-offs in terms of sum-rate, user outage, and energy efficiency, while disclosing the major power consuming elements of user DQN. Finally, crucial open research issues are identified for realizing energy-efficient AI at all network levels.