Discovering Lattice Reduction Strategies via Self-Play

The article presents DeltaStar, a novel lattice reduction strategy discovered through deep reinforcement learning, which outperforms the traditional Lenstra-Lenstra-Lovász (LLL) algorithm. DeltaStar is trained as a single-player Markov Decision Process using a deep residual network and an AlphaZero-style self-play approach, demonstrating significant efficiency by requiring fewer row operations than LLL. Notably, it generalizes effectively to unseen moduli and dimensions up to 32 without retraining, making it a valuable tool for practitioners in cryptography and computational number theory.