OK, so the probability of the model (y?) evaluated to -inf at this starting point, even though all the prior distributions were small negative integer numbers. Right?
Would a “soft” exponential distribution mitigate that problem? There is a hard zero for half the space. That would make the derivatives small. Would MCMC be happier if the distribution was not so binary? Perhaps add a low-amplitude, wide Gaussian to the Exponential.
Is that (adding a low-ampltiuyde Gaussian to the Exponential) something I can do in a CustomDist? Can I add two distributions? Or do I need to modify distributions.Exponential to return a small number instead of zero? What would you advise?
You also recommended tighter priors. That would keep the sampling closer to the good solutions, and reduce the chance of a -inf log probability? Is that the right model for the suggestion?
Thanks for the help.
- Malcolm