CRiSP: Reinforcement Learning for Quantum State Preparation

Researchers propose CRiSP (Clifford Reinforcement Learning agent for State Preparation), a framework that uses reinforcement learning to improve initialization in Variational Quantum Algorithms (VQAs). VQAs face optimization challenges like barren plateaus and local minima. CRiSP formulates discrete prefix selection as a sequential decision-making problem, employing Neural-Guided Monte Carlo Tree Search with a Transformer-based policy trained via self-play. It inserts learned Clifford gates before fixed parameterized rotations, enabling high-quality initial states through polynomial-time classical stabilizer simulation without altering the circuit architecture. The method aims to scale better than heuristic-based approaches in vast combinatorial search spaces.