基于替代函数及贝叶斯框架的1范数ELM算法

针对极端学习机 (Extreme learning machine, ELM)算法的不适定问题和模型规模控制问题,本文提出基于1范数正则项的改进型ELM算法. 通过在二次损失函数基础上引入1范数正则项以控制模型规模,改善ELM的泛化能力.此外,为简化1范数 正则化方法的求解过程,利用边际优化方法,构建适当的替代函数,以便于采用贝叶斯方法代替计算复杂的 交叉检验方法,并实现正则化参数的自适应估计.仿真结果表明,本文所提算法能够有效简化模型结构,并 保持较高的预测精度.

An Norm 1 Regularization Term ELM Algorithm Based on Surrogate Function and Bayesian Framework

Focusing on the ill-posed problem and the model scale control of ELM (Extreme learning machine), this paper proposes an improved ELM algorithm based on 1-norm regularization term. This is achieved by involving an 1-norm regularization term into the original square cost function, and it can be used to control the model scale and enhance the generalization capability. Furthermore, to simplify the solving process of the 1-norm regularization method, the bound optimization algorithm is employed and a suitable surrogate function is established. Based on the surrogate function, the Bayesian algorithm can be used to substitute the complicated cross validation method and estimate the regularization parameter adaptively. Simulation results illustrate that the proposed method can effectively simplify the model structure, while remaining acceptable prediction accurate.