基于人工免疫系统的RBF网络设计

RBF network designing based on artificial immune

由于传统的RBF网络学习方法存在诸多的不足,本文提出基于免疫机制的三级RBF网络学习方法:在第一级得到网络隐层节点数作为疫苗,不仅可自行构建网络,还降低了第二级搜索空间的复杂度;第二级利用人工免疫算法对解空间进行多点搜索,得到全局最优的隐层非线性参数;第三级采用最小二乘法确定网络输出层线性参数,极大地降低了第二级结构的维数,提高了算法效率。经典型Hermit多项式逼近实验验证了该方法训练得到的RBF网络性能优越。

In order to improve the traditional RBF learning strategy, a three-level RBF network learning algorithm based on immune system is proposed, which can calculates the number of the hidden-layer neurons in the first level as immune vaccine, the network can be established and adjusted by itself, and the complexity of search space in the second level can be reduced. The global optimum hidden-layer nonlinear parameters are searched for in the second level by parallel searching with artificial immune algorithm. The output-layer linear parameters are estimated in the third level with least square method, which makes the design dimension of the second level decreased and the algorithm efficiency improved. The experiment of Hermit polynomial approximation shows that the performance of the RBF network trained by the algorithm is superior.