Proximal Policy Optimization Reinforcement Learning Assisted Patch Antenna Design

To achieve efficient and low-cost optimization of electromagnetic (EM) response, a fixable design method for patch antenna based on proximal policy optimization (PPO) reinforcement learning (RL) algorithm is investigated in this paper. Specifically, the design parameters of the antenna are represented as a pixel matrix, allowing for a more flexible design space. For this character, the actor network optimizes each metal patch individually and expresses the results in probabilistic form, which ensures the ability to learn from optimization targets in a high degree of freedom design space. On the other hand, the critic network evaluates the design based on the target objectives and current performance, ensuring convergence towards the design goals. Two numerical case studies involving discrete variables are presented to verify the effectiveness of the proposed method.