Cobaya: chains stuck

Use of Healpix, camb, CLASS, cosmomc, compilers, etc.
Post Reply
Chen Heinrich
Posts: 5
Joined: October 06 2018
Affiliation: Caltech

Cobaya: chains stuck

Post by Chen Heinrich » October 13 2020

Hi Antony and Jesus,

I have a question about what the error message in Cobaya on chains being stuck really means in the MCMC run.

I have set the max_tries = 100, and followed advice on the documentation to set learn_proposal = False and/or using much smaller number in 'propose' than the expected 1D posterior. I get the error message that chains are stuck after N tries where N = max_tries * number of dimensions.

I also tried setting max_tries = 10^7 (effectively inf, which I realize is not recommended), but after running chains for about 0.1 million samples, the posterior of most parameters look Gaussian enough, the convergence criteria R_{-1} is however still infinity and not improving. How is this possible?

What is the criteria actually used to determine whether the chain is stuck? Is this a strong indication that the posteriors may be multimodal and I should be better off using a nested sampler? Please let me know your insights. Thank you!

Chen

Post Reply