Supplying initial values in matlabStan

[Daniel Lewis]

I'm pretty new to Stan - pushed through the learning process using the documentation and posts here (very helpful resource!), but I'm stuck on getting MatlabStan to accept user-supplied initial values. My model fits fine without supplied initial values, but when I try to fit with them, I get:

Warning: Stan seems to have exited badly.
> In StanFit/process_exit_failure (line 280)
  In StanFit/process_exit (line 170)
  In StanFit>@(src,evnt)process_exit(self,src,evnt)
  In processState/set.exitValue (line 30)
  In processManager.pollTimerStop (line 504)
  In timer/timercb (line 30)
  In timercb (line 13)
  In timer/start (line 39)
  In processManager/start (line 399)
  In StanModel/sampling (line 747)
  In stan (line 137)

I'm entering 

fit1 = stan( 'model_code',SV_code,'data',SV_data,'warmup',1000,'iter',10000,'chains',1,'init',inits)

where inits is given by

inits=struct('H21',.64,'H31',-2.11,'H12',-2.36,'H32',-1.69,'H13',-2.69,'H23',-.27)

As far as I can tell, this should be right, as according to the help file, init can be "a struct, with fields corresponding to parameters to be initialized". I assume I'm mistaken, and that's what's causing the error.

My model is below:

SV_code= {

'data {'

 ' int<lower=0> T;   // # time points (equally spaced)'

  'matrix[T,3] y;      // observed shocks'

'}'

'transformed data {'

'vector[3] zeros; //means are zeros'

'vector[T] yzeros;'

'yzeros = rep_vector(0, T);'

'zeros = rep_vector(0, 3);'

'}'

'parameters {'

 'corr_matrix[3] Omega;'

 'vector<lower=-1,upper=1>[3] phi;'

 'vector<lower=0>[3] sigma; '

 ' matrix[T,3] h_innov;             // std log volatility time t'

  'real H21;'

 ' real H31;'

 ' real H12;'

 ' real H32;'

 ' real H13;'

 ' real H23;'

'}'

'transformed parameters {'

 ' matrix[T,3] h;                 // log volatility at time t'

  'matrix[3,3] H;'

 'cov_matrix[3] Sigma; '

  'Sigma = quad_form_diag(Omega, sigma); '

 'H[1, 1] = 1; H[1, 2] = H12; H[1, 3] = H13; '

 'H[2, 1] = H21; H[2, 2] = 1; H[2, 3] = H23;'

 'H[3, 1] = H31; H[3, 2] = H32; H[3, 3] = 1; '

 ' h = h_innov;'

 ' h[1,1] = h[1,1] ./ sqrt(1 - phi[1] * phi[1]);'

 ' h[1,2] = h[1,2] ./ sqrt(1 - phi[2] * phi[2]);'

 ' h[1,3] = h[1,3] ./ sqrt(1 - phi[3] * phi[3]);'

 ' for (t in 2:T)'

 '   h[t] = h[t] + h[t-1] * diag_matrix(phi);'

'}'

'model {'

'vector[3] vol_t;'

 'Omega ~ lkj_corr(1); // LKJ prior on the correlation matrix ;'

 'sigma ~ cauchy(0, 2);'

  'for (t in 1:T){'

  'h_innov[t] ~ multi_normal(zeros,Sigma);'

  'vol_t=h[t]'' ;'

  'y[t] ~ multi_normal(zeros, H*diag_matrix(exp(vol_t))*H'');'

  '}'

  'phi ~ uniform(-1,1);'

  'H21 ~ normal(0,1);'

  'H31 ~ normal(0,1);'

  'H12 ~ normal(0,1);'

  'H32 ~ normal(0,1);'

  'H13 ~ normal(0,1);'

  'H23 ~ normal(0,1);'

'}'

};

Thanks so much!

Daniel

[Bob Carpenter]

I don't think Brian Lau, the MatlabStan dev, reads this list,
so you're probably better off filing an issue directly on
the MatlabStan repo:

https://github.com/brian-lau/MatlabStan/

Hopefully you won't need to supply your own inits---Stan is
pretty good at getting from diffuse inits to the typical
set.

- Bob

> --
> You received this message because you are subscribed to the Google Groups "Stan users mailing list" group.
> To unsubscribe from this group and stop receiving emails from it, send an email to stan-users+...@googlegroups.com.
> To post to this group, send email to stan-...@googlegroups.com.
> For more options, visit https://groups.google.com/d/optout.

[Daniel Lewis]

Thanks for the reply - I'll ask over there. I'm really happy with stan's inits, I'm just trying to explore a local minima I another algorithm is stuck in.

Daniel