基于显著性分析的神经网络混合修剪算法

Hybrid pruning algorithm for the neural network based on significance analysis

针对神经网络结构设计问题,提出了一种混合修剪算法。该算法首先利用合作型协同进化遗传算法和反向传播算法的不同优势,完成了对神经网络的结构和权值的初步调整;然后,在保证模型泛化能力的前提下,通过计算隐层神经元的显著性,借此修剪网络中显著性较小的神经元,进一步简化网络结构。最后,将给出的基于显著性分析的神经网络混合修剪算法用于股票市场的预测。仿真结果表明,该改进算法与其他优化算法相比,具有更好的泛化能力和更高的拟合精度。

This paper puts forward a kind of hybrid pruning algorithm for considering the problem of neural network structure design. Firstly, the algorithm uses the different advantages of cooperative coevolutionary genetic algorithm and back propagation algorithm to optimize the structure and weights of neural networks. Secondly, by calculating the significance of the hidden layer neurons, it prunes the network that is not significant, further simplifying the structure of the network without reducing the generalization ability of the model. Finally, the proposed hybrid pruning algorithm is used to forecast the stock market. The simulations showed that the improved algorithm has better generalization ability and higher fitting precision than other optimization algorithms.