For truss topology optimization problems, a hybrid method of reinforcement learning (RL) and graph embedding (GE) is developed. The GE method handles input data as a graph composed of nodes and edges, which is advantageous to grasp the structural property of trusses because the connectivity information of joints and members is explicitly maintained during the operations. The extracted features are utilized to estimate the performance change due to the removal of each member, and the accurate estimation is obtained using an RL method.
Owing to the mini-batch learning scheme using block matrices, the proposed method is capable of simultaneously handling a variety of trusses of different numbers of nodes and members. The proposed method is applied to topology optimization of trusses for volume minimization under stress constraints against multiple loading conditions.
During the training, a variety of support and load conditions are provided for an initial ground structure, and the removal of unnecessary members and training the model are implemented simultaneously. In numerical examples, the trained model shows a good performance for not only trained conditions but also unexperienced conditions without re-training. The trained model is expected to become a supporting tool that quickly feedbacks a near-optimal truss topology, which may enhance our design exploration towards better structural designs.