Flexible network reconfiguration is considered as an essential characteristic of active distribution network and it is helpful for accommodating distributed energy resources high-efficiently. However, many conventional mathematical programming methods can not tackle the problem of non-convex distribution network reconfiguration with DG. This paper presents a Discrete Learning Optimization Algorithm (DLOA) to solve this problem, which mainly includes three modules, learning optimization algorithm, discrete strategy, and topology analysis technique. It takes learning optimization algorithm as the core of program optimization, uses discrete strategy to determine the on/off status of distribution line, and applies topology analysis technique to analyze the structure of distribution network. The validity of DLOA is testified through 33-bus test system. Simulation results show that DLOA can effectively solve the highly non-convex distribution network reconfiguration with DG. This work is supported by Technology Project of Guangdong Power Grid Company (No. GDKJQQ20152029).