Aiming at the NP-hard characteristics of directional antenna network topology design under cluster UAV background, an elite strategy for deep reinforcement learning communication network topology generation algorithm is introduced with the requirements of high survivability, low power consumption and high stability of the network, which has the rewarding of invulnerability (3-connectivity), link quantity, link power consumption and stability. Compared with traditional DQN, elite experience pool verifies the acceleration training effect by effectively accelerating the convergence of the model and reducing the training time by more than three times. Rather than genetic algorithm, this algorithm separates the processes of use and training . When the network training is completed, the communication network topology can be calculated in real time with the needs of scene. In experimental stage, a 3-connected communication network topology with randomly given spatial location is designed which includes 6 nodes, 10 nodes, 24 nodes and 36 nodes . The experimental results has shown that this proposed algorithm has strong realtime and applicability, it can help network topology which has less than 36 nodes update in 183 ms so that meeting the realtime requirements of practical application.