多策略MRFO算法的卷积神经网络超参数优化

卷积神经网络的性能与超参数配置密切相关,然而最优超参数的选择耗时耗力. 为了提高超参数选择的效率,提出了一种基于多策略的蝠鲼觅食优化算法,一方面采用半数均匀初始化策略提升种群的多样性;另一方面,融合新权重因子更新策略和分裂策略,提升收敛速度和拟合精度. 根据实数编码策略将所提算法用于卷积神经网络的超参数优化研究中,用3种觅食方式进行迭代,以得到最优的超参数配置. 为了评估超参数优化的有效性,与卷积神经网络超参数优化算法在手写数字和CIFAR-10数据集上进行了对比实验,实验结果表明,所提算法可消耗较少的资源,并获得更高的准确率.

Hyperparameter Optimization of Convolutional Neural Network Based on Multi-Strategy MRFO Algorithm

The performance of convolutional neural network is closely related to the configuration of hyperparameters. However, the selection of optimal hyperparameters is time-consuming and labor-consuming. In order to improve the efficiency of hyperparameter selection, a multi-strategy manta ray foraging optimization algorithm is proposed. On the one hand, half uniform initialization strategy is adopted to improve population diversity. On the other hand, it combines new weight factor update strategy and splitting strategy to improve the convergence speed and fitting accuracy respectively. According to the real coding strategy, this algorithm is applied to the research of convolutional neural network hyperparameter optimization, which can be iterated according to three foraging methods to obtain the optimal hyperparameter configuration. In order to evaluate the effectiveness of hyperparameter optimization, the proposed algorithm is compared with the mainstream convolutional neural network hyperparameter optimization algorithms on mixed national institute of standards and technology and CIFAR-10 datasets. Experimental results show that the proposed algorithm achieves higher accuracy with less resources.