基于罚函数内点法的泄露积分型回声状态网的参数优化

发布时间：2018-05-11 13:59

本文选题：回声状态网 + 时间序列预测　；参考：《自动化学报》2017年07期

【摘要】：为了提升泄露积分型回声状态网(Leaky integrator echo state network,Leaky-ESN)的性能,提出利用罚函数内点法优化Leaky-ESN的全局参数,如泄漏率、内部连接权矩阵谱半径、输入比例因子等,这克服了通过反复试验法选取参数值而降低了Leaky-ESN模型的优越性和性能.Leaky-ESN的全局参数必须保障回声状态网满足回声状态特性,因此它们之间存在不等式约束条件.有学者提出利用随机梯度下降法来优化内部连接权矩阵谱半径、输入比例因子、泄露率三个全局参数,一定程度上提高了Leaky-ESN的逼近精度.然而,随机梯度下降法是解决无约束优化问题的基本算法,在利用随机梯度下降法优化参数时,没有考虑参数必须满足回声特性的约束条件(不等式约束条件),致使得到的参数值不是最优解.由于罚函数内点法可以求解具有不等式约束的最优化问题,应用范围广,收敛速度较快,具有很强的全局寻优能力.因此,本文提出利用罚函数内点法优化Leaky-ESN的全局参数,并以时间序列预测为例,检验优化后的Leaky-ESN的预测性能,仿真结果表明了本文提出方法的有效性.
[Abstract]:In order to improve the performance of leaky integrator echo state network (Leaky-ESNs), the penalty function interior point method is proposed to optimize the global parameters of Leaky-ESN, such as leakage rate, spectral radius of internal connection weight matrix, input scale factor, etc. This overcomes the superiority of the Leaky-ESN model and the global parameters of the Leaky-ESN model. The global parameters of the Echo state net must be guaranteed to satisfy the echo state characteristics, so there are inequality constraints between them. Some scholars have proposed that the stochastic gradient descent method can be used to optimize the spectral radius of the internal connection weight matrix, the input scale factor, and the leakage rate. To some extent, the approximation accuracy of Leaky-ESN is improved. However, the stochastic gradient descent method is the basic algorithm for solving unconstrained optimization problems. The parameter must satisfy the constraint condition of echo characteristic (inequality constraint condition is not considered), so the parameter value obtained is not the optimal solution. Because the penalty function interior point method can solve the optimization problem with inequality constraints, it has a wide range of applications, fast convergence speed and strong global optimization ability. Therefore, a penalty function interior point method is proposed to optimize the global parameters of Leaky-ESN. The prediction performance of the optimized Leaky-ESN is tested by using time series prediction as an example. The simulation results show the effectiveness of the proposed method.
【作者单位】：渤海大学新能源学院;渤海大学工学院;
【基金】：国家自然科学基金(61573072,21506014) 辽宁省自然科学基金(2014020143) 2011年辽宁省第一批科学技术计划项目(2011402001) 辽宁省教育厅科学技术研究项目(L2015008)资助~~
【分类号】：O211.61;TP183

【相似文献】