Sequential GP-UCB Bayesian Optimization for Deep Neural Network Fine-Tuning in Dissolved Oxygen Prediction

Abstract

Abstract Dissolved Oxygen (DO) is a key indicator of water quality, essential for sustaining aquatic ecosystems and human uses. Machine learning, particularly deep learning, is recognized as an effective approach for predicting DO levels by learning from data rather than requiring explicit human knowledge input. The effectiveness of deep learning models improves with fine-tuning of hyperparameters. Amongst hyperparameter tuning methods, Bayesian methods have gained particular interest for optimization. This study focuses on predicting DO levels in riverine environments using a Deep Neural Network model. The research employs a Gaussian Process Upper Confidence Bound (GP-UCB) Bayesian optimization technique to fine-tune hyperparameters, aiming for an optimal configuration. Comparative analysis is conducted between the optimized model and baseline model with default settings. Results indicate that the Bayesian-optimized model outperforms the baseline, particularly evident with moderately sized datasets. The findings underscore the pivotal role of Bayesian optimization in elevating model performance, exhibiting robust generalization capabilities while significantly reducing the need for manual parameter tuning. This successful application underscores a substantial methodological advancement in environmental management, particularly in predictive modelling for indicators of aquatic ecosystem health.