径向基概率神经网络结构的遗传优化

运用遗传算法(GA)来优化设计径向基概率神经网络(RBPNN)结构，优选了隐中心矢量和优化求取对应的核函数控制参数．提出的染色体编码方式，充分体现了所选隐中心矢量在模式样本空间中的数量及位置分布，同时还包含了相适应的棱函数控耕参数信息．新构造的适应度函数不仅有效地控制了网络输出的误差精度，而且还能够使得RBPNN结构优化趋于最简．将IRIS分类问题用于检验该算法的有效性并与ROLSA和MKM进行了比较研究，结果表明，GA的优化效率最高，而且GA优化后的RBPNN在推广能力方面也没有明显下降．

Genetic Optimization of the Structure of Radial Basis Probabilistic Neural Networks

The genetic algorithm is used to optimize the structure of the radial basis probabilistic neural networks (RBPNN), including selecting the hidden center vectors of the first hidden layer and determining the corresponding controlling parameters of the kernel function of RBPNN. The proposed genetic encoding method completely embodies the number and position distribution for the selected hidden center vectors in the pattern sample space, and involves information of the corresponding controlling parameters of the kernel function. The novel constructed fitting function can not only efficiently control the error accuracy of the RBPNN output, but also make the optimized RBPNN approach the most parsimonious structure. For a validity test, the proposed GA is used to solve IRIS classification problem and compared with ROLSA and MKM. The experimental results show that, the optimized performance of the proposed GA is the best among the three methods, with no obvious drop observed in the generalization performance of the optimized RBPNN by GA.