Session: DAC-25-01: AI-Driven Design Innovation

Paper Number: 115280

115280 - Teaching Ai to Design From Humans: A Comparison of Behavioral Cloning Architectures 

Reinforcement Learning (RL) has shown promise in creating agents with superhuman performance in various fields, including gaming and robotics. However, using RL methods to automate engineering design tasks is challenging due to the inability to generalize and the slow training process. Moreover, the most advanced RL algorithms require exploring millions of design states, which is infeasible for expensive physics models. Even the most advanced curiosity-based RL algorithms require exploring millions of design states, which is infeasible in engineering design tasks with expensive physics models. In contrast, data from a human-subject design study indicates that novice human designers can solve related engineering design tasks by exploring a few hundred design states. In this paper, we explore the performance of behavioral cloning, an imitation learning algorithm, for engineering design tasks to imitate the policies of a human designer from their decision data. We train a behavioral cloning agent on human design decision data collected in a controlled experiment. We then compare two popular sequence learning methods to behavioral cloning. Our results show that while sequence learning methods outperform behavioral cloning in predicting human design decisions as measured by mean absolute error, the behavioral cloning agent converges to a design solution by exploring relatively few design states.

Presenting Author: Joseph Thekinen University of Calgary

Presenting Author Biography: Dr. Joseph Thekinen is an Assistant Professor in the Department of Mechanical and Manufacturing Engineering at the University of Calgary. He received his Ph.D. from the School of Mechanical Engineering at Purdue University. He did his B.Tech and M.Tech from Indian Institute of Technology Kharagpur, with honors such as university silver medal (for topping academics in his department), best Masters thesis award, ABS scholarship, and J.P. Ghose Memorial award. His research focuses on enhancing human-AI partnership in decentralized sociotechnical systems using algorithmic game theory, machine learning, network science, and experimental techniques.

Authors:

Ghazal Boozarjomehry University of Calgary
Joseph Thekinen University of Calgary