Dueling-DQN Based Spectrum Sharing Between MIMO Radar and Cellular Networks

We investigate a two-tier distributed spectrum sharing framework between a multi-cell multi-user mobile broadband network (MBN) and a multiple-input multiple-output (MIMO) radar. While multi-agent reinforcement learning (RL) is used for transmit power allocation at MBN's base-stations to improve the quality of service of its users subject to the constraint of the probability of detection of the radar, the interference from the radar towards the MBN is mitigated via null-space based waveform projection. In the RL framework, the multiple cells in the MBN operate as agents, and the average signal-to-noise ratio value is the reward. Accordingly, we propose a deep RL network called dueling deep Q-network (DDQN) to enable co-existence by taking into account the physical layer parameters of the MBN and radar communication. The DDQN is compared to two other baseline RL algorithms, namely Q- learning and deep Q-network (DQN). Numerical results show that DDQN learns to obtain the best power allocation policies for distributed spectrum access without needing a centralized controller to control interference towards the radar. In particular, over time, the advantage network of DDQN allows the agent to take actions having higher advantage value, thus leading to faster convergence and a more stable spectrum sharing framework.