模块化神经网络的Bayes子网集结新算法研究

Novel integrated algorithm of modular neural network's sub-nets based on Bayesian learning

针对模块化神经网络的重要命题——子网动态集成问题,提出一种基于改进的Bayes学习的子网集结新方法.首先从处理复杂问题能力、计算开销、训练误差限等级的合理性、逼近正确率的构造等方面分析了已有方法的不足.既而提出相应策略,其核心在于采用了简洁、相关性小的子网生成方法;同时以误差作为依据提出新的逼近正确率指标以确定子网的动态集结权值.仿真实验对两种改进方法的测试误差进行了比较研究,结果表明了改进方法的有效性.

Aiming at the important issue of modular neural network （MNN） the dynamic integration of the sub-nets, a novel integrated algorithm based on the improved Bayesian learning is presented. Firstly, the drawbacks of the old algorithm are analyzed from four aspects-the processing ability for complex problems, computing cost, the rationality of trained error limit and the approximated accuracy. Then the corresponding strategy is presented whose key points are adopting a concise and lowly correlative sub-nets generating method, and then a new approximated accuracy index based on the error measure is presented to determine sub-nets＇ dynamic integration weights. The effectiveness of the above algorithms was demonstrated by simulation through the comparative research to two improved algorithms＇ test accuracy.