基于新型进化规划的异构神经网络集成算法

提出一种基于新型进化规划的异构神经网络集成算法。首先利用改进的进化规划，克服传统的进化规划及EPNet模型的不足，生成多个异构的最优网络，然后对异构网络进行组合求解。该集成算法充分利用了Bootstrap采样的天然特性实现了网络间的异构和差异度，同时又保证了单个成员网络的精度，克服了传统Bagging，Boosting算法中成员网络结构固定，缺乏个体精度的缺点。通过仿真实验证明，该方法较传统集成算法具有更好的泛化性能并减少了传统集成算法中的随机不确定因素。

An algorithm of Heterogeneous Neural Network Ensemble Based on New Evolutionary Programming

An algorithm of Heterogeneous Neural Network Ensemble Based on New Evolutionary is presented in this paper. At first, through the improving evolutionary programming, the shortcoming of traditional evolutionary programming and EPNet model can be overcome and heterogeneous neural networks can be generated, then these networks are integrated to get the result. This method make use of the advantage of Bootstrap sampling to realize the difference between the networks and guarantee the precision of the individual network. The simulation tests prove that this algorithm has higher generalization ability compare to traditional ensemble method and reduce the random element at the same time.