Development of an Artificial Intelligence System for Structural Design using Reinforcement Learning: Proof of Concept

摘要

Artificial intelligence (AI) has significantly progressed during the last several decades with the rapid advancement in computational capabilities. This advanced technology is currently being applied in various engineering fields, not just in computer science. Artificial neural network (ANN) is currently being widely used in mechanical/structural problems since ANN can be used as a surrogate model for modeling the physical system. However, these applications do not employ the highest AI level, in which the ANN can judge by itself, as is the case of advanced AI algorithms such as an autonomous driving car or a walking robot. Reinforcement Learning (RL) is an approach to machine learning that mimics human behavior, like how human beings solve a problem based on their experience. A human problem-solving process can be improved by earning a positive reward from good experiences (results). Also, the RL algorithm can determine which actions will cause a worse outcome and provide negative feedback. Therefore, such an algorithm can be applied to solve structural design problems where the engineers can efficiently resolve the issues and bring correct results through trial and error. In this study, an AI system with the RL algorithm is developed to design optimized truss structures (with continuous and discrete cross-section choices) under a set of given constraints. Also, we proposed a unique reward function system to consider the constraints in structural design problems. From a set of two examples, we confirmed that the proposed AI system could design truss structures and also evolve as it gains experience. Therefore, it is possible to develop an AI system that can learn from experience and design the structure by itself without little human intervention.