Deep Reinforcement Learning based Mode Selection and Power Allocation Scheme for Convergent Power-Efficient Network

The ever-growing data traffic has posed a great challenge to celluar networks, incuring resource scarcity and high energy consumption problems. Broadcast and multicast are believed to alleviate the pressure of the network because they of the advantage of serving multiple users with the common spectrum resources. However, power-efficient mode selection and power allocation according to the users' distribution under different interference scenarios are urgent tasks to be achieved. In this paper, a deep reinforcement network(DRN) based scheme is proposed to determine the delivery mode and transmitting power of BSs with a given transmission requirement. With the objective of minimizing the power consumed to serve each user while assuring a high service rate, we formulate the mode selection and power management issue as a mixed integer nonlinear programing problem (MINLP). To solve this problem, a reinforcement learning based scheme is proposed, which uses the neural network to capture the geographical distribution of users and BSs, and then produce the decent solution without accurate CSI information. The policy gradient scheme is employed to achieve a proper parameter update in an reinforcement manner, which eventually converges to promising solutions. Simulation results demonstrate that over 20% power efficiency gain can be achieved by the proposed scheme at the median and high intercell interference(ICI) level while assuring a high service rate.