Reinforcement Learning-based Pauli Network circuit synthesis

There is growing interest on applying AI to optimization of quantum circuits. On previous work, we showed how by applying RL to circuit synthesis and routing we were able to achieve a good balance between optimality and computational effort. We also presented a general framework to apply RL to similar problems that arise on quantum circuit transpiling.

Here, we build upon this framework to do architecture-aware synthesis of general Pauli Networks. We encode the Pauli Network synthesis problem as a sequential decision problem, and use the RL framework to train models that can do synthesis of arbitrary Pauli Networks and follow constraints imposed by physical devices (like qubit connectivity).

We train models for synthesis up to 7 qubits, and use these for peephole optimization of general quantum circuits, achieving over 30% 2-qubit gate reduction on top of Qiskit SDK transpiler for some circuits.