Stan++/Stan3 Preliminary Design

[martinmodrak]

Thanks for the response. I understand that this is hard and I am in the comfortable position of having an opinion but not having any responsibility for the decision :-)

However, I think my main point still got lost in translation, giving it another try :-)

@Bob_Carpenter

So if you wanted, you could do this:

model gaussian_process() {
   data int<lower = 0> N;
   data vector[N] times;
   param real length;
   param real sigma;   
   data vector[N] y ~ multi_normal(0, cov_exp_quad(times, length, sigma));
}

data int<lower = 0> N;
data vector[N] times;
param real length;
param real sigma;   
model gp = gaussian_process()

Now it looks almost like an existing model in terms of the order of declarations.

But how does the compiler determine that times in the main model maps to gp.times (the submodel?). That is the magic that worries me. I’d rather have to write something like

model gp = gaussian_process(N, times, length, sigma, y);

or even with named parameters to avoid mismatches in long param list.

model gp = gaussian_process(
             N = N, 
             times = times, 
             length = length, 
             sigma = sigma, 
             y = y
           );

This way it is obvious what gets passed where and I can be precise what is shared and what is different in each submodel. For example there is now a clear distinction between:

model gp1 = gaussian_process(N, times, length, sigma, y1);
model gp2 = gaussian_process(N, times, length, sigma, y2);

and

model gp1 = gaussian_process(N, times1, length1, sigma, y1);
model gp2 = gaussian_process(M, times2, length2, sigma, y2);

Hope that clarifies my concern :-)