Мультиагентне навчання з підкріпленням для оптимізації квантових схем

• Date: 2025
• Main Authors: Kyrylov, Ivan, Sinitsyn, Igor
• Format: Article
• Language: Ukrainian
• Published: V.M. Glushkov Institute of Cybernetics of NAS of Ukraine 2025
• Subjects: мінімізація кількості вентилів; квантова компіляція; квантові обчислення
• Online Access: https://jais.net.ua/index.php/files/article/view/531
• Journal Title: Problems of Control and Informatics

The article presents an approach to quantum circuit optimization based on Multi-Agent Reinforcement Learning (MARL), which integrates the MAPPO algorithm with Graph Neural Networks (GNNs). The relevance of the research stems from the need to reduce gate counts and circuit depth in quantum compilation to enhance noise resilience and execution efficiency on current quantum devices. Unlike traditional single-agent approaches, the proposed architecture, MAQCO (Multi-Agent Quantum Circuit Optimizer), enables task distribution among agents, facilitating better scalability, coordination, and consideration of the local context of the circuit. The optimization process is implemented using equivalence-preserving transformations from the Quartz system, allowing simplification of circuits without altering their functionality. Experimental evaluation on benchmark circuits using Nam and IBM gate sets demonstrates that MAQCO achieves superior or comparable results to state-of-the-art optimizers such as Quarl and Quartz, particularly in reducing total gate count, CNOT (CX) gates, and circuit depth. Future research directions include scaling MAQCO to larger circuits, integrating ECC and ZX-based transformations, and adapting the method for hardware-aware optimization on distributed quantum architectures, taking into account topology constraints and communication costs.