IPC 2022 featured speaker
Dr Xianzhong Zhao is a Professor in Civil Engineering at Tongji University and the Head of the Intelligent Design & Construction (IDC) research group. He is a distinguished professor of the Changjiang Scholars Program and was the Dean of the College of Civil Engineering at Tongji University from 2016 to 2021.
His research areas include steel structures, progressive collapse of structures, as well as intelligent design and construction following his work on digital design as a Research Associate (RA) at the University of Cambridge.
Professor Zhao has published more than 100 academic papers, has been authorized many granted patents and served as the principal investigator (PI) of over 30 research and consulting programs both in China and abroad.
Professor Zhao leads the establishment of the functional platform for R&D and transformation of intelligent construction at Tongji University, an International Journal of AI in Civil Engineering and the first undergraduate degree program ‘Intelligent Construction’ in China. He is the Secretary-General of the Steering Committee on Civil Engineering Teaching in Higher Education, and the Secretary-General of the Appraisal Group of Civil Engineering for the Academic Degrees Committee of the State Council.
An MDP-based reinforcement learning algorithm
for intelligent design
of spatial structures
Computers have been successfully applied to structural analysis, design, and drawing in civil engineering. However, how to “utilize computer to simulate human brain” for intelligent design is still a great challenge.
Alongside with AlphaGo in 2016, Monte Carlo Tree Search (MCTS)-based algorithms have made history for beating professional players in Go games by solving Markov Decision Processes (MDPs). It is a landmark event in the field of reinforcement learning (RL), showing that machines can surpass the vast majority of people in intellectual activities. Inspired by this success, structural intelligent design for truss structures may benefit from MCTS by splitting the design process into an MDP model since the solution space of the optimal truss layout problem is huge in terms of size, shape, and topology optimization.
This study presents an MDP-based RL model for generating the optimal truss layout. In this MDP model, the solution space is greatly expanded through three sequential action sets. The reward function of the MDP can give feedback to actions according to both geometric stability and structural simulation. To find the optimal sequential actions in solution space, a MCTS-based algorithm named AlphaTruss is proposed in this study, which can give the best decision in each design step by searching and learning through the MCTS technique. Other reinforcement learning methods are also welcome to solve this MDP model for structural intelligent design.