Use latin hypercube sampler to initialize GP optimizations

dflemin3 / approxposterior

A Python package for approximate Bayesian inference and optimization using Gaussian processes

https://dflemin3.github.io/approxposterior/

MIT License

41 stars 9 forks source link

Use latin hypercube sampler to initialize GP optimizations #46

Closed jlustigy closed 4 years ago

jlustigy commented 5 years ago

This could help to ensure that the global minimum is actually found, and that any potential local minima are included in the cross-validation calculation (see #41).

dflemin3 commented 4 years ago

Added GP hyperpriors and dramatically improved GP hyperparameter optimization in #55 and earlier commits. With these changes, maximizing the GP marginal likelihood seems to work well in general for approxposterior's use cases.