基于近似结构风险的ELM隐层节点数优化

隐层节点数是影响极端学习机(ELM)泛化性能的关键参数,针对传统的ELM隐层节点数确定算法中优化过程复杂、容易过学习或陷入局部最优的问题,提出结构风险最小化-极端学习机(SRM-ELM)算法。通过分析VC维与隐层节点数量之间的关联,对VC信任函数进行近似改进,使其为凹函数,并结合经验风险重构近似的SRM。在此基础上,将粒子群优化的位置值直接作为ELM的隐层节点数,利用粒子群算法最小化结构风险函数获得极端学习机的隐层节点数,作为最优节点数。使用6组UCI数据和胶囊缺陷数据进行仿真验证,结果表明,该算法能获得极端学习机的最优节点数,并具有更好的泛化能力。

Optimization for Hidden Nodes Number of ELM Based on Approximate Structure Risk

The number of hidden nodes is a critical factor for the generalization of Extreme Learning Machine（ELM）.There exists complex optimization process,over learning or traps in local optimum in traditional algorithm of calculating the number of hidden layer of ELM. Aiming at the problems,Structural Risk Minimization（SRM）-ELM is proposed.Combining empirical risk with VC confidence,this paper proposes a novel algorithm to automatically obtain the best one to guarantee good generalization. On this basis,the Particle Swarm Optimization（PSO）position value is directly treated as ELM hidden layer nodes,which employs the PSO in the optimizing process with Structural Risk Minimization（SRM）principle. The optimal number of hidden nodes is reasonable correspond to6 cases. Simulation results show that the algorithm can obtain the extreme learning machine optimal nodes and better generalization ability.