Deep Reinforcement Learning based Resource Allocation in NOMA

NOMA is a novel channel accessing strategy that delivers high throughput and fairness among various users by multiplexing many users across the same frequency resource. In order to guarantee the user's fairness, minimum data rate maximization, also referred to as the max-min approach is adopted. Apparently, transmission power optimization is employed to accomplish the max-min. However, the scalability of the number of users leads the optimization to a non-convex optimization problem. Consequently, the Dueling Double Deep Q Learning(Dueling DDQL) technique, a subclass of Reinforcement Learning is proposed to solve such problem. The Deep Q-Network is used by the DDQL approach in learning the actions that are best to do to maximize user power coefficients. The Markov Decision Process (MDP) model is essential to the DDQL method's effectiveness since it trains the DQN on choosing better actions. The dueling DDQL converges to the target value for 92% of the test cases. The proposed method is compared with the benchmark algorithms and it is illustrated that the proposed algorithm outperforms those comparative algorithms.