Optimization of Single-Layer Reticulate Shell Assembly Sequence Using Deep Reinforcement Learning Graph Embedding Method

This study explores reinforcement learning algorithms combined with graph embedding methods to optimize the assembly sequence of complex single-layer reticulate shells. To minimize the number of temporary support brackets during installation, the structural assembly process is modeled using the inverse dismantling process. The remaining members of the structure at each iteration step are scored, and the one with the highest score for removal is selected. Next, this study trains an effective intelligent agent to assemble the structure. The proposed method can be used to design several types of latticed shells. The trained intelligent model can complete the assembly sequence design of the mesh shell without requiring any other data except for previous structural information. To verify the feasibility of the novel method, it is compared with the empirical approach used in the traditional assembly sequence design process. The feasibility of the new method is demonstrated. It is indicated that the novel method can obtain the optimal solution accurately and efficiently. In addition, it has more innovative choices for installation sequences than the conventional technique. It has enormous potential and application in the civil engineering field.