A Reinforcement Learning Protocol for Solving the Maximum Independent Set Problem on Neutral-Atom Processors

We present a reinforcement learning (RL) framework, which autonomously explores analog control schedules for pulsed laser sequences, to solve the Maximum Independent Set (MIS) problem using neutral-atom quantum processors. Our protocol initializes atom registers from graph instances, builds time-dependent driving field Hamiltonians, and extracts the probability of success from a physical or simulated backend. This technique utilizes the MIS probability as a reward signal, and implements policy-gradient methods to optimize laser amplitude and detuning schedules under Rydberg blockade constraints. We also incorporate pre-training using heuristic baselines with fast training and periodic validation, to further optimize performance. Our results demonstrate that learned schedules reliably match, and even outperform, hand-engineered ramps for benchmark MIS instances and generalize across different graph families; which we also compare to state-of-the-art graph-agnostic methods. The result is a modular, reproducible toolkit that unifies graph generation, encoding, simulation, evaluation, and visualization, thereby lowering the barrier to reliable “RL for quantum control” studies, and facilitating the rapid iteration from simulation to experiment.