(362o) Improved Performance of Artificial Neural Networks Via Hyperparameter Optimization and Data Augmentation for a Small Number of Data Sets

Hyper-parameter optimization and a sufficient amount of data with quality are crucial for a reliable prediction of artificial neural networks (ANNs). However, collecting enough data to be learned by ANN models requires a lot of time and effort, and may be quasi-impossible in reality. In this study, the ammonia emission rate consisting of eleven input variables and two outputs (maximum ammonia loss, N_max and time to reach half of N_max, K_m) was considered, which had 83 data sets. The data were pretreated using maximum normalization, one-hot coding for categorical input data, and generative adversarial network (GAN) for data augmentation. Hyper-parameters were optimized through Gaussian process (GP). The ANN model improved the prediction performance, reducing the mean absolute error by 38.02% in K_m and 56.02% in N_max. The correlation coefficient (R²) was significantly improved compared to the results from the literature. The proposed method is useful for regression problems with a small number of data sets.