Bandwidth Enhancement of Pixelated Patch Antennas Based on Localized Reinforcement Learning

This paper employs a reinforcement learning algorithm to improve bandwidth for a pixelated patch antenna. Differing from traditional methods that depend on manual tuning and extensive computations, this approach autonomously discerns the relationships between numerically characterized pixelated patch antennas and their S-parameters. Specifically, the implementation of a double deep Q-network (DDQN) and prioritized experience replay allows the model to dynamically adjust learning processes based on a structured exploration of state and action spaces. In this case, pixelated patch antenna configurations are efficiently optimized for maximal bandwidth performance. The learning model systematically tests and refines antenna designs by iteratively adjusting the presence or absence of metal at each pixel, thereby achieving a significant bandwidth increase to 34%. This method not only optimizes the bandwidth efficiently but also suggests promising applications in complex antenna designs.