PSIS.jl implements the Pareto smoothed importance sampling (PSIS) algorithm from Vehtari et al, 2021. Given a set of importance weights used in some estimator, PSIS both improves the reliability of the estimates by smoothing the importance weights and acts as a diagnostic of the reliability of the estimates.
Vehtari A, Simpson D, Gelman A, Yao Y, Gabry J. (2021). Pareto smoothed importance sampling. arXiv:1507.02646v7 [stat.CO]
The following packages also implement PSIS:
- PSIS: Matlab and Python reference implementations
- loo
- arviz
- ParetoSmooth.jl