Interface roadmap - last draft before ratification vote

[avehtari]

@seantalts

We’ll get rid of thin, jitter, kappa, gamma, and t0, as we haven’t seen any use cases for these parameters and think maintenance and pedagogy are easier without them.

Here is an example which would kill Bob’s computer (and probably mine, too) without thin

Pseudo-extended MCMC

Thanks. Here’s the Stan code: transformed data { real r = 10; real A = 6; real omega = 6; } parameters { vector[2] x; } model { target += -0.5 * (hypot(x[1], x[2]) + r + A * cos(omega * atan2(x[2], x[1]))); } This is probably another good case for convergence diagnoistics. Here’s the RStan call: fit <- sampling(model, control = list(stepsize = 0.01, adapt_delta = 0.999, max_treedepth=11), chains = 4, iter = 50000) There was still one iteration tha…

thin is rarely needed, but there are cases where if thin is removed, things get very difficult.

@seantalts

Here are the various quantities that will have this functionality:

• effective_sample_size
• rhat
• treedepth
• energy
• quantile
• mean
• variance

How about advi diagnostic? (there is currently experimental code in R for advi in rstan and rstanarm, and for optimizing in some cases in rstanarm)