講演情報
[9a-F212-7]Hybrid Quantum-Classical Architecture Search with Multi-Objective Reward Shaping for Intermediate-Complexity Quantum Machine Learning
〇Yang Xiao1, Reo Saito1, Juncheng Wang1, Koki Awaya1, Takumi Kanezashi1, Haruya Nagata1, Jun-ichi Shirakashi1, Tetsuo Shibuya2, Hiroshi Imai2 (1.Tokyo Univ. Agr. & Tech., 2.Univ. Tokyo)
キーワード：
Quantum Architecture Search、Quantum Machine Learning、Reinforcement Learning
Quantum architecture search (QAS) is a promising approach for automating variational quantum circuit design, yet existing reinforcement learning-based methods often suffer from sparse rewards and unstable convergence on nontrivial tasks. We propose a hybrid RL-QAS framework that incorporates a trainable classical neural network layer to enhance expressive power without increasing quantum resources. Additionally, a phase-adaptive multi-objective reward function dynamically balances classification performance, circuit compactness, and training stability throughout the search process. Experiments on the Wine dataset show that the proposed method achieves stable convergence and automatically discovers compact circuit architectures, reaching a peak accuracy of 0.97, while the baseline RL-QAS fails to converge. The results demonstrate the effectiveness of combining classical neural augmentation and adaptive reward shaping for scalable quantum machine learning.