Learning-Assisted Energy Minimization for MEC Systems With Noncompletely Overlapping NOMA

This article investigates the long-term energy minimization of mobile edge computing (MEC) systems with multiuser noncompletely overlapping nonorthogonal multiple access (NOMA) when time varying channel is considered. Different from the existing work with multiuser noncompletely overlapping NOMA that adopted the delay-based successive interference cancellation (SIC) decoding order, we adopt the channel gain-based SIC decoding order, which is widely used and has been proved to be more energy-efficient. We formulate the problem as a joint transmit power, transmission time, and computing resource optimization problem with multiple deeply coupled variables. To ensure the real-time decision-making, the considered problem is characterized as a deep reinforcement learning (DRL) problem by introducing penalty mechanism into the immediate reward function when the constraints are violated, and we propose a soft actor-critic (SAC) algorithm to solve it. Simulations show that the proposed algorithm can significantly improve the system's performance, and, compared with the noncompletely overlapping NOMA mechanism with delay-based SIC decoding order, the completely overlapping NOMA mechanism and time division multiple access (TDMA) mechanism, the noncompletely overlapping NOMA mechanism with channel gain-based SIC decoding order can achieve an average energy consumption reduction of 32.4%, 49.9%, and 74.4%, respectively.