A Compensation for Elevated Sidelobe of Radiation Pattern of Antenna Array Caused by Amplitude and Phase Discretization Based on Deep Reinforcement Learning

In our previous research, we proposed a multi-layer perceptron method for synthesizing radiation patterns with complex requirements. However, this method has a limitation in that it requires continuous amplitude and phase excitations. The continuous excitations may not be practical in most applications, and when the amplitude and phase of the antenna element are discontinuous due to the resolution of the electronic device during the optimization process, this method may fail. In order to solve this problem, a combined method is proposed in this paper, utilizing the multi-layer perceptron network to optimize the radiation patterns of array antenna with continuous amplitude and phase, and introducing a deep reinforcement learning method to compensate for the elevated sidelobes in the radiation patterns caused by amplitude and phase discretization in post-processing. Specifically, the continuous amplitude and phase values obtained through the multi-layer perceptron network are rounded to approximate discrete results based on the resolution, serving as prior training experience for the deep reinforcement learning model. The compensation problem of elevated sidelobes caused by amplitude and phase discretization is then formulated as an optimization model, and a deep reinforcement learning model is constructed accordingly, with discrete excitations acting as the agent for exploration and search. The agent is trained using the Deep Q-learning network as the basic framework, combined with Double DQN technology and Dueling DQN technology to efficiently search for the best compensatory effect. The reward is carefully designed to incentivize the agent to search for discrete excitations with the optimal compensatory effect. A simulation experiment is conducted on a 50-element hemispheric conformal antenna array, demonstrating the effectiveness of the combined method.