Deep deterministic policy gradient and graph convolutional network for bracing direction optimization of grid shells

Abstract

Read online

In this paper, we propose a method for bracing direction optimization of grid shells using a Deep Deterministic Policy Gradient (DDPG) and Graph Convolutional Network (GCN). DDPG allows simultaneous adjustment of variables during the optimization process, and GCN allows the DDPG agent to receive data representing the whole structure to determine its actions. The structure is interpreted as a graph where nodes, element properties, and internal forces are represented by the node feature matrix, adjacency matrices, and weighted adjacency matrices. DDPG agent is trained to optimize the bracing directions. The trained agent can find sub-optimal solutions with moderately small computational cost compared to the genetic algorithm. The trained agent can also be applied to structures with different sizes and boundary conditions without retraining. Therefore, when various types of braced grid shells have to be considered in the design process, the proposed method can significantly reduce computational cost for structural analysis.

Keywords

• bracing direction optimization
• reinforcement learning
• deep deterministic policy gradient
• graph convolutional network
• grid shell structures