论文部分内容阅读
目的:比较多元线性回归、逐步多元回归(SMR)、BP网络(BPNN)、基于遗传算法的BPNN(GA-BPNN)、支持向量回归(SVR)和最小二乘SVR(LS-SVR)6种方法用于均匀设计(UD)优化的效果。方法:以某UD实验数据为研究载体,采用平均绝对百分比误差等为评价指标,对6种方法进行了精度、泛化能力及优化效率等的评价。结果:SMR方法简洁,需时少;BPNN稳健性较差,GA-BPNN性能较BPNN大为提高;SVR准确度高,泛化能力强,优于其他方法。结论:SMR、GA-BPNN、SVR和LS-SVR 4法均可较好地用于该载体的优化;对于变量较少、非线性不复杂而精度要求不苛刻时的UD优化,宜用SMR;而对于变量较多、非线性较为复杂而精度要求较高时,SVR较宜,其次GA-BPNN和LS-SVM。
Objectives: To compare 6 methods of multiple linear regression, stepwise multiple regression (SMR), BP network (BPNN), genetic algorithm based BPNN (GA-BPNN), support vector regression (SVR) and least squares SVR For uniform design (UD) optimization of the effect. Methods: Taking a UD experimental data as the research carrier, using the average absolute percentage error as the evaluation index, the accuracy, generalization ability and optimization efficiency of the six methods were evaluated. Results: The SMR method is simple and requires less time; the BPNN is less robust, the performance of GA-BPNN is much better than that of BPNN; SVR is more accurate and more generalized than other methods. CONCLUSIONS: SMR, GA-BPNN, SVR and LS-SVR 4 methods are both suitable for the optimization of this vector. SMR is suitable for UD optimization with less variables, non-linearity and less strict accuracy requirements. However, for more variables, more complex nonlinearities and higher accuracy, SVR is preferable, followed by GA-BPNN and LS-SVM.