Optimizing Reservoir Computing with Genetic Algorithm for High-Dimensional SARS-CoV-2 Hospitalization Forecasting: Impacts of Genetic Algorithm Hyperparameters on Feature Selection and Reservoir Computing Hyperparameter Tuning
Résumé
This study focuses on forecasting SARS-CoV-2 hospitalizations 14 days ahead at Bordeaux University Hospital using a high-dimensional dataset with 409 predictors and 586 observations, combining public data and electronic health records. Previous research showed that integrating reservoir computing (RC) with genetic algorithm (GA) for hyperparameter optimization and feature selection outperformed state-of-the-art methods. However, the behavior of RC-GA under high-dimensional conditions is not well understood. This work examines the impact of GA hyperparameters (GA-HP), specifically the mutation probability of feature selection and the extent of mutation on a critical RC hyperparameter (RC-HP), the leaking rate. GA-HP significantly influence feature selection and RC-HP optimization. Higher mutation rates led to a more diverse set of selected features and fewer total features, resulting in increased mean absolute error (MAE) on the training set but comparable MAE on the test set, suggesting reduced overfitting. Conversely, lower mutation rates of categorical genes and leaking rates correlated with slightly poorer performance, indicating potential lack of exploration during RC-HP selection. Notably, a bimodal Ferté et al. convergence of the leaking rate was observed, with lower leaking rates enhancing training performance but slightly diminishing test performance. The RC-GA approach demonstrated robust behavior with higher leaking rates, possibly due to increased regularization effects from higher ridge values. Optimizing RC-GA in a high-dimensional setting remains challenging. Current practices using the median forecast of the top 40 RC-HP sets might not be optimal. Enhancing GA for high-dimensional contexts and exploring complementary RC-HP sets instead of selecting the top 40 might further improve forecasting accuracy.
Origine | Fichiers produits par l'(les) auteur(s) |
---|