Deep reinforcement learning based multicast mode selection for SFN

The dramatic increase in data traffic poses a great challenge to the cellular network. To alleviate the pressure of the network and improve the spectrum efficiency, Multimedia Broadcast and Multicast Service (MBMS) is introduced to enable the point-to-multipoint transmission in the cellular network. To enhance the received signal strength (RSS) of cell edge users, the Multicast/Broadcast Single Frequency Network (MBSFN), which allows combining signals from different cells, is adopted. However, neighbor cells may incur the interference and include more users in bad conditions, which leads to lower transmission efficiency. As a result, an appropriate selection of the multicast mode for each cell is needed. In this paper, a deep reinforcement network (DRN) was proposed to obtain near-optimum solutions for the SFN area formation. We first design the input of the neural network to characterize the geographic locations of users. The output of the neural net is set to be a probability matrix, which parametrizes the mode selection status. A reinforcement learning framework based on the policy gradient scheme was adopted to bypass the acquisition of optimal labels. By training with different network layouts, DRN can produce near-optimum solutions for throughput maximization. The average increase of DRN above the traditional approach is also revealed by numerical simulation.