sel_inf_dev to lognormal

[grantdadams]

will make penalties between slp and inf devs on similar scale

[grantdadams]

Or use some other scale invariant transformation so that the SD can be shared across slp and inf devs (code). I dislike the hard-coded 4 * SD....

            jnll_comp(5, flt) -= dnorm(sel_inf_dev(0, flt, sex, yr), Type(0.0), sel_dev_sd(flt), true);
            jnll_comp(5, flt) -= dnorm(ln_sel_slp_dev(0, flt, sex, yr), Type(0.0), 4 * sel_dev_sd(flt), true);

This is a carry over from the ADMB pollock assessment (see code) and maybe there is an option for better parameterization across models @Cole-Monnahan-NOAA ?

Note that Rceattle also does not assume the asymptote is positive:

            Type inf   = sel_inf(0, flt, sex) + sel_inf_dev(0, flt, sex, yr);
            Type val = 1.0 / (1.0 + exp(-slope * (x_val - inf)));

Maybe an alternative would be:

            Type inf   = **exp**(sel_inf(0, flt, sex) + sel_inf_dev(0, flt, sex, yr));
            Type val = 1.0 / (1.0 + exp(-slope * (x_val - inf)));

[Cole-Monnahan-NOAA]

Agree we should change it. How about just making them their own parameters (scalars) and we can map them off and fix at the data input vals?

[jimianelli]

I think it might be worthwhile to move away from rigid TV selex parameters and instead use parametric forms (if wanting them so bad) as the backbone/init and then use AR1 or AR2-like devs, if that makes sense? Here's an example:

get_selectivity_prior <- function(rho_y, rho_a, log_sigma, par_log_sel_fya) {
  "[<-" <- ADoverload("[<-")
  n_sel <- 7
  lp_sel <- numeric(n_sel)
  for (f in seq_len(n_sel)) {
    scale <- exp(log_sigma[f]) / sqrt(1 - rho_y[f]^2) / sqrt(1 - rho_a[f]^2) # define 2d scale
    f1 <- function(x) dautoreg(x, phi = rho_y[f], log = TRUE) # year
    f2 <- function(x) dautoreg(x, phi = rho_a[f], log = TRUE) # age
    lp_sel[f] <- -dseparable(f1, f2)(par_log_sel_fya[[f]], scale = scale)
   }
  return(lp_sel)
}

[Cole-Monnahan-NOAA]

We (ie.., Grant) tried that in 2023 and didn't work well. Need to revisit it though.