Rocky Raccoon: Automated quantum circuit optimization using graph-based deep reinforcement learning

Efficient compilation of quantum algorithms is crucial to the success of near-term and fault-tolerant quantum information processing devices. In general, circuit synthesis and gate decomposition results in quantum circuits with suboptimal gate count and circuit depth. Quantum circuit optimization aims to address this limitation by finding and applying strategies to simplify quantum circuits and reduce the amount of resources required to implement quantum algorithms. We propose a framework for quantum circuit optimization by formulating it as a Markov decision process, and use graph-based deep reinforcement learning to train an agent to optimize quantum circuits by applying local and simple rewrite rules. The agent in our framework is realized by a graph neural network which operates on a directed acyclic graph representation of a quantum circuit. The actions of the agent are formulated as simple rewrite rules applied to pairs of adjacent quantum gates. We'll discuss our ongoing work on benchmarking the framework and how the proposed agent may be used to discover complex circuit optimization strategies in an automated fashion by sequentially composing simple rewrite rules.