Splitting schemes are numerical integrators for Hamiltonian problems that may advantageously replace the { St\"ormer-Verlet} method within Hamiltonian Monte Carlo (HMC) methodology. However, HMC performance is very sensitive to the step size parameter; in this paper we propose a new method in the one-parameter family of second-order of splitting procedures that uses a well-fitting parameter that nullifies the expectation of the energy error for univariate and multivariate Gaussian distributions, taken as a problem-guide for more realistic situations; we also provide a new algorithm that through an adaptive choice of the $b$ parameter and the step-size ensures high sampling performance of HMC. For similar methods introduced in recent literature, by using the proposed step size selection, the splitting integration within HMC method never rejects a sample when applied to univariate and multivariate Gaussian distributions. For more general non Gaussian target distributions the proposed approach exceeds the principal especially when the adaptive choice is used. The effectiveness of the proposed is firstly tested on some benchmarks examples taken from literature. Then, we conduct experiments by considering as target distribution, the Log-Gaussian Cox process and Bayesian Logistic Regression.
翻译:在汉密尔顿-蒙特-蒙特-卡洛(HMC)方法中,HMC的性能对步数参数非常敏感;在本文件中,我们提议在单数分解程序第二阶次的单数组中采用一种新的方法,该二阶分解程序使用一个非常合适的参数,使对单向和多变戈西亚分布的能源错误的期望完全无效,在比较现实的情况下,这种参数可作为问题指南;我们还提供一种新的算法,通过调适性选择美元参数和分级大小确保HMC的高采样性能。对于最近文献中引入的类似方法,我们采用拟议的步数选择,HMC的分解整合方法在对单向和多变制高斯分布应用时绝不拒绝样本。对于更一般的非高斯目标分布,特别是在采用适应性选择时,拟议的办法超出了本位。提议的算法的有效性首先在从文献中对某些基准示例进行测试,然后我们考虑将Log-Cressirvicrial。