Ho to increase the number of effective samples using Metropolis

The problem of Metropolis is that in high dimension the random walk is nearly impossible to explore the posterior space properly in finite time. For a complex system like PDE, I would suggest you to study the posts by @aseyboldt (see more details here Using pymc3 NUTS sampler for an "external model" having its derivatives with respect to its parameters). In many cases, gradient could be approximated which allow you to use NUTS.

1 Like