|
Abstract:
|
Improving designs in physical experiments takes years of practice and expert knowledge. To expedite the process, scientists develop representative mathematical models and run computer simulations to discover possible solutions. However, most algorithms suffer issues such as dependency on large datasets to train the algorithms as well as computational infeasibility as models grow in size and complexity. Additionally, modern data has become more complex like curves and images, where many existing methods only apply to traditional experimental design with scalar inputs and outputs. We address these problems by combining functional data analysis with humans-in-the-loop (HITL). First, functional data analysis represents and visualizes several types of data structures from scalars to manifolds. Second, incorporating the experts in tandem with statisticians for HITL reduces the parameter search space for learning algorithms, provides insight into messy or missing data to avoid implementation issues, and guides the next set of simulations to improve. We present real-world applications that demonstrate our proposed approach.
|
