Reinforcement Learning and Graph Embedding for Binary Truss Topology Optimization Under Stress and Displacement Constraints

This paper addresses a combined method of reinforcement learning and graph embedding for binary topology optimization of trusses to minimize total structural volume under stress and displacement constraints. Although conventional deep learning methods owe their success to a convolutional neural network that is capable of capturing higher level latent information from pixels, the convolution is difficult to apply to discrete structures due to their irregular connectivity. Instead, a method based on graph embedding is proposed here to extract the features of bar members. This way, all the members have a feature vector with the same size representing their neighbor information such as connectivity and force flows from the loaded nodes to the supports. The features are used to implement reinforcement learning where an action taker called agent is trained to sequentially eliminate unnecessary members from Level-1 ground structure, where all neighboring nodes are connected by members. The trained agent is capable of finding sub-optimal solutions at a low computational cost, and it is reusable to other trusses with different geometry, topology, and boundary conditions.