A Portable Auto-Tuning Framework for Quantum Compilation Optimization Based on D3QN

This paper proposes a portable quantum compilation optimization framework based on D3QN (Double Dueling Deep Q-Network) to automate quantum circuit tuning. Addressing challenges in adapting fixed strategies to diverse algorithms, the framework integrates reinforcement learning (RL) with a quantum compiler environment. The RL agent dynamically selects optimization passes, guided by a reward function balancing gate count, circuit depth, and hardware constraints. Quantum circuits are encoded as DAGs, with D3QN estimating Q-values and standardizing action-space encoding, state representation, and constraint validation. Experiments on the MQTBench dataset (430 circuits across 18 algorithms) show the framework outperforms QCOR-O3 and random strategies, achieving an average 8.7% gate reduction. Implemented on QCOR's MLIR-based compiler, the modular design ensures portability. Results validate generalization across algorithms, though performance varies for underrepresented circuits. This work establishes a standardized, RL-driven framework for quantum compilation optimization in NISQ systems.