Hard-instance learning for quantum adiabatic prime factorization

Prime factorization is a difficult problem with classical computing, whose exponential hardness the foundation of Rivest-Shamir-Adleman (RSA) cryptography. With programmable quantum devices, adiabatic computing has been proposed as plausible approach to solve prime factorization, having promising advantage over computing. Here, we find there are certain hard instances that consistently intractable for both simulated annealing and un-configured (AQC). Aiming at an automated architecture optimal configuration apply deep reinforcement learning (RL) method configure AQC algorithm. By setting success probability worst-case reward RL, show performance on dramatically improved by RL configuration. The also becomes more evenly distributed different instances, meaning configured stable compared case. Through technique transfer learning, prominent evidence framework scalable -- trained five qubits remains working efficiently nine minimal amount additional training cost.