Human-in-the-Loop Bayesian Optimization for Artificial Intelligence-Guided Design

Abstract

Abstract. To address large-scale problems such as ocean cleanup and reforestation, engineers need to develop novel solutions that can effectively cover a wide area and adapt to a variety of situations. To design a complex system that can solve these problems, such as a fleet of devices, it is necessary to have methods that can guide engineers toward optimal designs. However, determining the optimal engineering specifications is a difficult task because there is a complex relationship between the individual device and how the system performs as a whole. This work develops a method to find these specifications using an artificial intelligence (AI)-guided design method based on human-in-the-loop Bayesian optimization. By representing design tradeoffs using machine learning, the method can effectively guide engineers by suggesting new design targets. We validate this method with a user study where participants build small prototypes that represent forest-replanting devices in a simulated fleet, with the goal of maximizing the fleet's effectiveness. The guided group followed the specifications given to them by the AI guide, while the unguided group chose their own specifications. While the mean effectiveness was similar between the groups, the guided group had higher baseline effectiveness than the unguided group. Additionally, there was no significant reduction in the variety of the guided group's designs. This shows that AI-guided design can consistently help engineers find optimal solutions without negative effects on design exploration.