Abstract
The emergence of quantum reinforcement learning (QRL) is propelled byadvancements in quantum computing (QC) and machine learning (ML), particularlythrough quantum neural networks (QNN) built on variational quantum circuits(VQC). These advancements have proven successful in addressing sequentialdecision-making tasks. However, constructing effective QRL models demandssignificant expertise due to challenges in designing quantum circuitarchitectures, including data encoding and parameterized circuits, whichprofoundly influence model performance. In this paper, we propose addressingthis challenge with differentiable quantum architecture search (DiffQAS),enabling trainable circuit parameters and structure weights usinggradient-based optimization. Furthermore, we enhance training efficiencythrough asynchronous reinforcement learning (RL) methods facilitating paralleltraining. Through numerical simulations, we demonstrate that our proposedDiffQAS-QRL approach achieves performance comparable to manually-craftedcircuit architectures across considered environments, showcasing stabilityacross diverse scenarios. This methodology offers a pathway for designing QRLmodels without extensive quantum knowledge, ensuring robust performance andfostering broader application of QRL.