层次模块化神经网络新方法研究

A Novel Algorithm for Hierarchical Modular Neural Network

多个神经网络组成的系统有两种工作方式:其一是通过训练多个神经网络,以协同或竞争的方式来构建学习系统解决同一问题;其二则是将复杂问题分解成若干简单问题而后进行处理。应用分而治之的思想提出了一种层次模块化神经网络新方法———三层结构的模块化神经网络HMNN(HierarchicalModularNeuralNetwork)模型,在提高算法性能方面有一定优势,对比实验研究也表明该算法有效地提高了系统的泛化能力和算法稳定性。

A system based on multiple neural networks usually has two ways of working. Firstly, many neural networks work cooperatively or competitively to solve a problem. Secondly, based on the idea of modularity, each neural network works for the subdivision of a complex object. A new architecture of multiple neural networks based on modularity is presented, named HMNN(Hierachical Modular Neural Network). The empirical study shows that, compared with some traditional neural network ensemble and modular neural network algorithms, HMNN has strong generalization and stability.