Improving Generalization Performance of Structural Optimization by Reinforcement Learning

Abstract

It can be said that it is an important problem for engineers and the construction industry to make general public understand as much as possible about the structural engineering which is difficult to understand. To deal with this problem, we thought that interactive operation would be effective in order to understand the essence of the structural engineering. Therefore, we have developed a structural optimization application using a touch panel. In addition, in previous study, we proposed a method using Reinforcement Learning (RL) for optimization logic with the goal of more flexible optimization, and could show the possibility that RL could be applied to structural optimization. However, some problems remained in terms of generalization performance because of the biased structural conditions of the learning environment and the method of Deep Q Network (DQN). In this study, for the purpose of improving generalization performance, we extended the learning environment and introduced Prioritized Experience Replay to DQN. As a result, it was confirmed that both generalization performance and optimization performance improved compared to the results of previous study.