基于和声搜索算法的极限学习机网络优化

Optimization of extreme learning machine network based on harmony search algorithm

极限学习机(ELM)因其运算速度快、误差小等优点而得到广泛的应用,但由于随机给定输入权值和阈值可能导致隐含层节点无效,因此,ELM通常需要增加隐含层节点数来提高预测精度,从而导致网络泛化能力不佳。为了解决上述问题,提出一种和声搜索算法的极限学习机网络(HS-ELM),采用和声搜索算法不断调整ELM输入权值和隐含层阈值矩阵选取最优以达到优化网络的目的。最后通过两种复杂度不同的非线性函数拟合加以验证。结果表明,传统ELM网络平均预测误差为0.31×10-3%和1.6%,HS-ELM的平均预测误差为0.01×10-3%和0.4%。证明和声搜索算法优化后的ELM网络在同等情况下所需的隐含层节点数和预测精度均优于传统ELM网络的。

Extreme learning machine(ELM)is widely used because of its fast operation speed and small error ability.However,random reference input weights and thresholds may lead to hidden layer nodes invalid.Therefore,ELM usually need to increase the number of hidden layer nodes to improve the prediction accuracy which result in poor generalization ability of network.In order to solve the above problems,an algorithm for the extreme learning machine of the harmony search algorithm is proposed by using harmony search algorithm to adjusting the ELM input weight and the hidden layer threshold matrix.Finally,two different of complexity nonlinear functions are tested and verified.The results show that the average prediction error of the traditional ELM network is 0.31×10-3%and 1.6%while the average prediction error of HS-ELM is 0.01×10-3%and 0.4%.It is proved that the number of hidden layer nodes is less while prediction accuracy is higher of ELM networks optimized by harmony search algorithm than those of traditional ELM networks under the same conditions.