Aiming at preventing the network performance deterioration caused by fixed link topology connections in optically switched networks, this paper proposes and validates an interconnection and control mechanism that can automatically optimize and reconfigure network topology according to the network traffic volume. This mechanism combines deep reinforcement learning with software defined network technology. Based on the real-time network performance parameters collected by the software-defined network control plane and the performance feedback of the network simulation system, the continuous training of the topology structure can be realized in the model of deep reinforcement learning. The automatic network topology reconfiguration is achieved with improved network performance. The experimental results show that for a given traffic intensity, the model trained by deep reinforcement learning can learn the topology reconstruction strategy and realize automatic selection. The optimal network topology structure can be achieved in one step, with reduced average network delay and average packet loss.