Hyperparameter Optimization of Multi-layer Perceptron-Based Predictors

Hyperparameter tuning, often viewed as a complex process due to the vast search space it involves, typically requires expert oversight. Tools like Keras Tuner and AutoML simplify this task by automating parts of the process. However, the multiplicity of hyperparameters can increase costs and extend computation times, a crucial aspect in applications where quick and effective responses are essential. In this paper, we introduce a two-phase hyperparameter optimization (HPO) method, specifically designed for univariate time series prediction. This method, evaluated in scenarios with both synthetic and real time series, demonstrates how effectively reducing the search space significantly enhances accuracy, offering an efficient solution in demanding environments.