Toward Quantum Reinforcement Learning for Automated Neural Architecture Optimization in Vision-Based Structural Health Monitoring

Abstract

Neural Architecture Search (NAS) involves exploring large and combinatorial search spaces, for which numerous approaches have been proposed, including reinforcement learning, evolutionary algorithms, Bayesian optimization, and gradient-based methods among others. Despite these advances, efficient exploration remains a key challenge, particularly as the search space scales combinatorially. Motivated by prior work based on Q-learning strategies, which are limited in large spaces, this work investigates a Quantum Reinforcement Learning (QRL) framework based on probabilistic policy optimization. We propose a QRL approach where a parameterized quantum circuit represents a policy over candidate architectures, and is trained to maximize the expected reward defined by validation accuracy. This method is compared against two baselines: Random Search and a Classical Reinforcement Learning approach using policy gradient with a Bernoulli distribution. All methods are evaluated under identical budgets using the NATS-Bench benchmark, enabling efficient and controlled comparison without full model training. The results demonstrate that both QRL and Classical RL learn meaningful search policies, consistently outperforming Random Search. QRL achieves competitive performance with respect to Classical RL, confirming its ability to effectively explore the architecture space. We further analyze the impact of key hyperparameters to identify practical trade-offs between accuracy, model complexity, and computational cost. Overall, the results demonstrate that QRL is a viable approach for discrete architecture search, capable of learning effective policies in large combinatorial spaces, while indicating clear directions for improving its efficiency and scalability.