Optimum design of space steel frames based on decomposition and reconstruction of plane frames using reinforcement learning

Abstract

In practical design of building structure, the cross-sectional dimensions of the members are selected from a prescribed standard list. Machine learning (ML) is potentially effective for this problem; however, preparing a large dataset in advance is difficult. Even in this situation, Reinforcement Learning (RL) is applicable, can assist to find a better solution, and is expected to reduce computational cost when compared without the RL agents. We propose a new minimum volume design of a building steel frame based on a decomposition and reconstruction framework using a RL agent. In the method, i) the design problem of a space frame is decomposed into that of plane frames, ii) the plane frames are optimized by standard nonlinear programming, iii) and the plane frames are assembled to a space frame by the RL agent. The advantages of the methods are presented through a numerical example.