Using Deep Learning to Simulate Multi-Disciplinary Design Teams

Human subject experiments are often used in research efforts to understand human behavior in design. However, such research is often time-consuming, expensive, and limited in scope due to the need to experimentally control specific variables. This work develops an initial digital simulation of team-based multidisciplinary design, where the actions of individual team members are simulated using deep learning models trained on historical human design trends. The main benefit of this work is to simulate design session events and interactions without human participants, developing a complimentary method to rapidly perform digital team-based experiments. This research merges the benefits of purely data-driven modeling with minimal assumptions about process, along with the strengths of agent-based modeling in which it is possible to tailor agent behavior. Initial results show that the simulated design team sessions are able to replicate trends and distributions compared to human-based team sessions, but run approximately 21 times faster than equivalent human subject studies. The multi-disciplinary design problem currently simulated is loosely coupled, in the sense that agent behaviors can be modeled in isolation of other agents and yet replicate the behavior of the ensemble. Future work will extend the agents to sense and respond behaviors that can be used to model tightly coupled problems, and truly evaluate team formulations.