As a research hotspot in the field of machine learning, ensemble learning improved the prediction accuracy of the final model by constructing and combining multiple basic models. In recent years, many experts and scholars are committed to combining deep networks with ensemble learning to improve the accuracy of neural network models in various scenarios and tasks. But not all neural networks are suitable for participating in the construction of ensemble models. Deep networks with ensemble learning require that the single neural network involved in the integration has high accuracy and great discrepancy with other networks. In the initial stage of deep networks with ensemble learning, the process of generating sets of candidate deep networks is first required. After studying an existing multiobjective deep belief networks ensemble (MODBNE) method, the Gaussian random field model is used as a pre-screening strategy in the process of generating the candidate deep network sets. Individuals with great potential for improvement are selected for fitness function evaluation so that a large number of neural network models with higher accuracy and the larger discrepancy between networks can be easily obtained, which effectively improves the quality of the solution and reduces the time consumed in training the neural networks.

集成学习作为机器学习领域的研究热点，通过构建和组合多个基本模型来提高最终模型的预测精度。近年来，许多专家学者致力于将深度网络与集成学习相结合，以提高神经网络模型在各种场景和任务中的准确性。但并不是所有的神经网络都适合参与集成模型的构建。具有集成学习的深度网络要求参与集成的单个神经网络具有较高的准确性，并且与其他网络有很大的差异。在集成学习深度网络的初始阶段，首先需要生成候选深度网络集的过程。在研究了现有的多目标深度信念网络集成（MODBNE）方法后，在生成候选深度网络集的过程中，采用高斯随机场模型作为预筛选策略。选择具有较大改进潜力的个体进行适应度函数评估，从而可以轻松获得大量精度更高、网络间差异较大的神经网络模型，有效提高解的质量，减少训练所消耗的时间神经网络。