Analog self-interference cancellation for full-duplex communication based on deep reinforcement learning

Abstract

Full-duplex (FD) communication systems are expected to be extensively used in wireless communications, however, their performance is severely limited due to the self-interference (SI). Traditional analog self-interference cancellation (ASIC) methods generally do not consider estimating the delay of the SI channel, thereby requiring a large number of taps to capture channel details. In order to effectively eliminate SI in situations with limited resources or space, we propose two novel ASIC schemes based on deep reinforcement learning (DRL), named Multi-Deep Q Network (Multi-DQN) scheme and DQN-Deep Deterministic Policy Gradient (DQN-DDPG) scheme. Specifically, for the Multi-DQN scheme, we use multiple DQN units to estimate the delay and attenuation of the SI channel discretely, which can effectively reduce the number of taps required for ASIC. To overcome the loss of discretization, the DQN-DDPG scheme utilizes DQN and DDPG units to estimate the delay and continuous attenuation of the SI channel, respectively. Simulation results indicate that both proposed schemes achieve a similar performance to the multi-tap methods with fewer taps. Additionally, the effectiveness of both schemes is verified across various scenarios, encompassing system configurations, hyperparameters, and channel changes.