An Antenna Optimization Framework Based on Deep Reinforcement Learning
Fengling Peng, Xing Chen
Abstract
A hybrid strategy-based general antenna optimization framework is proposed. This framework integrates differential evolution (DE), customized decision trees, and deep Q-networks. It employs an artificial intelligence (AI) model based on supervised learning to train an AI model based on deep reinforcement learning (DRL), thereby enabling efficient exploration of antenna design solutions. The framework is divided into three phases: first, an improved DE is executed to achieve a preliminary global exploration of antenna design solutions. Second, the samples collected in the first phase are used to construct a decision tree, which provides rapid reward computation for subsequent reinforcement learning. Third, the optimal solution found in the first phase is used as the starting point for DRL of antenna design solutions, achieving a refined search of the solution space. Moreover, in the third phase, the decision tree is updated to achieve customization, thereby improving its accuracy in evaluating antenna design solutions. Experimental results show that this hybrid strategy-based antenna optimization framework can obtain superior antenna design solutions with fewer simulation iterations.