Reinforcement learning-based architecture search for quantum machine learning

Quantum machine learning (QML) leverages the large Hilbert space provided by quantum computing for data encoding, typically realized by parameterized quantum circuits (PQCs). While classical machine learning deals extensively with problem-specific model design, QML models mainly use hardwareefficient and heuristic circuit designs for PQCs. This work presents a novel approach employing the reinforcement learning algorithm MuZero to generate problem-specific PQCs to improve the QML performance. Diverging from previous search algorithms, we adopt a layered circuit design to significantly reduce the search space. Furthermore, we utilize cross-validation scoring to train the reinforcement learning algorithm, rewarding the discovery of high-performing circuits. In benchmarks we compare our tailored circuits with reference circuits from the literature, randomly generated circuits, and circuits generated by genetic algorithms. Our findings underscore the efficacy of problem-tailored encoding circuits in enhancing QML model performance.