论文部分内容阅读
高斯过程回归(GPR)学习机有着容易实现、超参数自适应获取及预测输出具有概率意义等优点。通常采用共轭梯度法获取GPR超参数,但其存在优化效果对初值依赖性太强,迭代次数难以确定,易陷入局部最优的缺点。改用粒子群优化(PSO)算法进行最优超参数搜索,形成粒子群-高斯过程回归耦合算法(PSO-GPR)。将该算法引入三峡永久船闸高边坡、卧龙寺新滑坡、链子崖滑坡3个不同的典型滑坡变形时序分析中,对每个滑坡分别采用稳态核及一种新式神经网络(NN)、平方指数(SE)、有理二次型(RQ)3种单一核函数进行外推预报测试。工程应用表明,基于3种不同单一核函数的粒子群-高斯过程回归算法(PSO-GPR)均能完全适应不同滑坡时序分析,其中以NN核函数外推预测效果最佳,平均相对误差分别为6.37%、7.62%、1.07%,从而改善了在进行不同滑坡变形时序分析时采用单一核函数的核机器外推能力存在较大差异性的问题,提高了单一核函数对不同数据类型的兼容性。
Gaussian process regression (GPR) learning machine has the advantages of easy realization, adaptive acquisition of hyperparameters and predictive output with probability significance. Usually, the conjugate gradient method is used to obtain the hyperparameter of GPR. However, its optimization effect is too dependent on the initial value, the number of iterations is hard to be determined, and it is easy to fall into the local optimum. The Particle Swarm Optimization (PSO) algorithm is used to search the optimal hyperparameters to form Particle Swarm Optimization - Gaussian Process Regression Coupling Algorithm (PSO-GPR). The algorithm was introduced into three different typical landslide deformation timing analysis of the permanent ship lock, the Wolongsi new landslide and the Zhiziya landslide. The steady state kernel and a new type of neural network (NN) were used for each landslide. The square Exponential (SE), rational quadratic (RQ) three kinds of single-kernel function extrapolation prediction test. The engineering application shows that PSO-GPR based on three kinds of single kernel functions can all adapt to different landslide time series analysis. Among them, NN kernel function extrapolation has the best prediction effect and the average relative errors are 6.37%, 7.62% and 1.07%, respectively, so as to improve the difference of extrapolation capability of nuclear machines adopting single kernel function in different landslide sequence analysis and improve the compatibility of single kernel function for different data types .