Soft-constrained reinforcement learning for antenna optimization with feasibility prescreening

Journal article, 2026

This work proposes a soft-constrained reinforcement learning (SC-RL) framework integrating feasibility prescreening and trust region mechanisms to address physical constraints in high-dimensional parameter space. By embedding penalty rules into the reward function, the framework transforms physical constraints into learnable optimization signals while leveraging proximal policy optimization (PPO) to stabilize policy updates. By establishing a unified abstraction where constraints are defined as mathematical penalty signals, the framework exhibits generalizability across varying antenna structures and radiation mechanisms. Experimental validation on a broadband antipodal linearly tapered slot antenna (ALTSA) and a wideband filtering patch antenna (FPA) demonstrates superior optimization performance and a favorable convergence rate compared to baseline methods. The ALTSA optimized by the proposed framework achieves a 21.1% impedance bandwidth, 6.90% axial ratio bandwidth, and 9.17 dBic peak gain, exceeding the original targets of 16%, 5%, and 8.8 dBic, respectively. While the optimized FPA attains a 23.1% impedance bandwidth and 9.96 dBi peak gain, surpassing its goals of 17% and 9.7 dBi. Furthermore, the convergence speed of this framework is 21.7% and 31.1% faster than standard PPO in the two cases, while outperforming traditional algorithms which are prone to either premature convergence or slower stabilization.