NBDO: An Autoencoder-Based Algorithm for Efficient Optimal Experimental Design in High-Dimensional Settings

Optimal experimental design is essential for obtaining precise statistical inference with limited resources, yet existing algorithms such as the Coordinate Exchange (CE) method can become computationally prohibitive in high-dimensional settings. We introduce the NeuroBayes Design Optimizer (NBDO), a novel algorithm that combines autoencoders with Bayesian Optimization to construct efficient experimental designs in reduced latent spaces. By compressing the design search space, NBDO enables fast exploration of complex design problems while directly optimizing an A-optimality criterion. We evaluate NBDO on two representative classes of models: second-order response surface models and scalar-on-function linear models. Across a range of scenarios, NBDO produces designs with efficiencies above 97–99% while achieving substantial reductions in runtime compared with both CE and its implementation in JMP. These results demonstrate that neural network–based methods can offer a scalable alternative to traditional search algorithms, particularly in settings with functional inputs or many experimental factors.