混合Kriging代理模型的高维参数估计优化算法

Hybrid Kriging surrogate model optimization algorithm for high-dimension parameter estimation

基于Kriging代理模型的优化算法对于解决函数计算昂贵的优化问题非常有效,但并不适用于高维参数的优化.针对该问题,提出了一个混合Kriging代理模型和多种优化技术的算法.该算法在Kriging模型选择新样例点时使用单维参数独立优化以克服维度灾难并提高收敛速度,同时基于已构建的Kriging代理模型信息提出一种新的动态坐标扰动策略,并将该策略用于高维参数优化以得到更好的目标函数值.为了保证不丢失全局最优解,在使用一般期望提高加点策略作为选点原则时,在期望函数的多个峰值同时选点.为了验证算法的有效性,将该算法应用于具有41维参数的人类白细胞代谢网络参数估计问题.实验结果表明,在有限的迭代次数下,该算法能产生较小的目标函数值,以及和实验拟合较好的参数估计结果.

Kriging surrogate-based model optimization algorithm is an effective algorithm in solving optimization problems with expensive computation.However,it is not feasible to deal with the parameter estimation of high-dimension.Aiming at this problem,a new optimization algorithm,i.e. hybrid Kriging surrogate model and other optimization technologies for high-dimension parameter estimation is proposed.In this algorithm,single parameter estimation is adopted for Kriging model infill sampling to conquer the curse of dimensionality and improve the convergence rate.Meanwhile, based on the information of the built Kriging surrogate model a new dynamic coordinate disturbance strategy is presented and used for high-dimension parameter estimation to refine objective values.To avoid missing the global optimal value,multi-modal searching based on sampling fill criterion.i.e. generalized expected improvement is introduced.Its effectiveness is verified by estimating parameters of a human polymorphonuclear leukocyte metabolic network with 41 dimension parameters.The experimental results show that the algorithm can produce better parameters＇; estimation results agreeing with experimental data and small objective values under limited number of iterations.