Reduced Routing Cost in Compiling Quantum Programs with Transformers and Tree Search

The success of quantum computing relies on an effective usage of the underlying Quantum Processing Units (QPUs). Compiling a logical quantum circuit tailored to the specific architecture of a physical QPU is a necessary step to run quantum circuits. This research presents a novel approach to quantum circuit compilation using Reinforcement Learning (RL). While traditional methods using solvers or heuristics produce correct compilation results, they suffer from three drawbacks: (1) their search takes a long time to run, (2) heuristics produce sub-par results, and (3) they require domain expertise to redesign to tailor to different QPUs.

Our RL-based model automatically generates compilation “experiences” to learn from to find optimized compilation policies. The RL-based method improves the three shortcomings of traditional methods. First, an RL agent simply does model inference during compilation with a significantly faster runtime. Second, an RL agent automatically learns a good strategy given arbitrary QPU architectures and target benchmarks to provide better performance. Third, RL replaces domain expertise development time by automatic model training time, hence eliminating a huge bottleneck in quantum system design.