Exploring Optimal Model for Machine Learning by Differential Evolution

Abstract

In the big data era, the data scientist or business analyst own business data can apply machine learning algorithm to train inference models easily since the application of big data attracts more and more attention and the technique of training models can be obtained easier and cheaper than ever. The focus of big data applications has gradually shifted from model training to the prediction and inference. For enterprise application scenarios, selecting the most precise model among many trained models has become a significant topic of research. Though ensemble methods have been proposed to discover best model by multiple training phase, studies of exploring best combination within multiple modes are still few. Finding the appropriate parameters to configure different machine learning models is an NP-hard problem that needs metaheuristic algorithm to solve. This study proposes a differential evolution algorithm to integrate multiple trained machine learning models into a hybrid model. For experiment, the regression model is taken as an example and the differential evolution algorithm is compared with the ant colony optimization algorithm in this paper. Three benchmark datasets are employed to examine, and the results discovered that the differential evolution algorithm outperforms ant colony optimization.