A Deep Reinforcement Approach for Energy-Efficient Resource Assignment in Cooperative NOMA-Enhanced Cellular Networks
Yanyan Guo, Xiao-Long Tan, Yun Gao, Jing Yang, Zhi-Chao Rui
Abstract
In this paper, an energy efficiency (EE) maximization problem of cooperative non-orthogonal multiple access (CNOMA) network is proposed to jointly determine the user pairing, subchannel assignment, and power control scheme. We decompose it into two steps: In the first step, the optimal closed-form expressions of the power control problem are derived. Based on these, the EE optimization of whole system is formulated as a self-play Go game with the maximum EE as the winner, by constructing a virtual Go board with rows and columns representing indices of users and subchannels, respectively. The each move is to select a position on the Go board (i.e., select a user that has not been assigned to any channel and then assign a channel to it), until all users are assigned on the subchannels. Then, a deep Monte Carlo tree search (MCTS) model is proposed, where a MCTS guided by a neural network simulates multiple possible trajectories to search each move by evaluating its achievable EE reward, while the neural network is trained by the training data generated from the searching of MCTS to predict move selections and also the winner of games. The simulation results show that the proposed method is superior to a variety of conventional schemes in terms of EE in negligible computational time.