Hybrid Quantum-Classical Reinforcement Learning in Patch Antenna Design

This paper presents a hybrid quantum-classical approach for the optimization of geometrical parameters of a circular patch antenna at 2.4 GHz for wireless applications. The information gathered during electromagnetic (EM) software simulations conducted in the initial design phase is repurposed as training data for constructing a Deep Neural Network (DNN) surrogate model. It serves as a substitute for resource-intensive EM simulations, thereby expediting the training process of reinforcement learning (RL). This acceleration facilitates the optimization process of the electromagnetic design. We also propose to integrate a Quantum Neural Network (QNN) composed of quantum circuits with the classical neural network to create a more expressive model. Moreover, the feasibility of the proposed optimization approach is investigated by comparing its results with those achieved through the traditional Deep Q-Network (DQN) RL algorithm. The experimental results from the hybrid model show its better training convergence while featuring 17 times smaller model complexity compared to regular DQN with only classical neural networks, despite significant time required for model training.