MULTI-AGENT REINFORCEMENT LEARNING FOR OPTIMAL DESIGN OF 3D-STEEL FRAMES AS ASSEMBLY OF 2D-FRAMES

Existing machine learning methods are difficult to apply to complex and large buildings with many design variables, and the design conditions are simplified in the formulation of optimal design problem. This paper presents a method using reinforcement learning of multi-agent for the problem of minimizing material volume of 3D-steel frames. In the learning process, two agents select the member and its cross-sectional dimension, respectively, with simple neural networks with small number of input features of local information. The numerical examples show that the method exploiting the two agents can optimize cross-sectional dimensions more efficiently than simulated annealing and local search.