Neural Network for Prediction of Functional Data

• Activity Number: 28 - Computation, Design, and Quality Assurance of Physical Science and Engineering Applications
• Type: Contributed
• Date/Time: Sunday, August 8, 2021 : 1:30 PM to 3:20 PM
• Sponsor: Quality and Productivity Section
• Author(s): Weiru Han* and Lu Lu and Jiangfeng Zhou
• Companies: Department of Mathematics & Statistics, University of South Florida and University of South Florida and Department of Physics, University of South Florida
• Keywords: multi-layer perceptron; functional principal component analysis; factorial design experiment; evolutionary algorithm; multiple objective optimization

Abstract

Neural networks are popular for classification and prediction involving more input than output variables. However, they have not been broadly used for predicting functional output. Motivated from a physical science study which involves predicting the spectrum curve from metamaterial optical device with different geometrical parameter settings, we developed neural network models that utilize the multi-layer perceptron for predicting the functional output. To optimize the network architecture, first a screening design was used to explore the impact of model hyperparameters such as the numbers of layers and nodes and the choices of activation functions. Then a two-phased genetic algorithm was developed to search the optimal structures to simultaneously minimize the training and prediction errors as well as the model complexity. The first phase uses a discrete non-dominated sorting genetic algorithm to seek top-ranking combinations of the number of nodes, activation functions and optimization methods. The second phases uses a continuous search focusing on optimizing the number of nodes. The proposed method will be illustrated with the optical device optimization example.