Accelerating Method of Evolutionary Ensemble Learning Based on Gaussian Random Field

Abstract

In recent years, research on ensemble learning of neural networks is very popular. As a research hotspot in the field of machine learning, ensemble learning methods can effectively improve the accuracy and generalization of deep network models, but not all neural networks are suitable for participating in the construction of ensemble model. Deep network ensemble learning requires a single neural network participating in the ensemble to have a high accuracy rate, and there is a large difference between the networks. In the initial stage of deep network ensemble learning, the generation process of the candidate deep network set is first required. In this article, a multi-objective evolutionary ensemble model is improved, and an evolutionary ensemble learning acceleration method based on Gaussian random field is added before the evaluation of fitness function which can screen individuals with great potential for improvement in the evaluation of fitness function during the generation of candidate deep network sets, thereby effectively improving the quality of the solution and reduce the time spent training neural networks. This pre-screening strategy is applied to the solution of the multi-objective differential evolution algorithm, which can conveniently obtain a large number of neural network models with high accuracy and large network differences. And this strategy speeds up the solution process of multi-target algorithm.