R devel - Feb 2005 - calling optif0 in a C function

Dear All,

I am trying to use the function optif0 (in main/uncmin.c) from the
latest R distribution. The reason is that I have a quite complicated
likelihood function which is coded in C, and I would like to optimize it 
directly.

To see how this works, I have tried with a very simple example: 
optimizing the likelihood of a sample using an exponential distribution. 
I have tried several solutions but none worked. I paste below the 
functions that come the closest to what should work. The compilation is 
fine and the call from R too. It seems that the call to optif0 does not 
do anything whereas everything else works.

It seems that I miss something simple... Any suggestion will be welcome. 
I use R 2.0.1 and GCC 3.3.4.

Best regards,

Emmanuel Paradis
=======================================================#include <R.h>
#include <Rmath.h>
#include <R_ext/Applic.h>
#include "ape.h"

/* Here is the content of ape.h: */
typedef struct {
   int *n;
   double *x;
} TITI;

static void fcn_expo(int, double *, double *, TITI *);
/* end of ape.h */


void lik_expo(int *n, double *x, double *l, double *loglik)
{
/* computes the likelihood */
   int i;
   *loglik = 0;
   for (i = 0; i < *n; i++)
     *loglik += log(*l) - *l * x[i];
}

static void fcn_expo(int np, double *p, double *sol, TITI *D)
{
/* computes the deviance to be minimized */
   double loglik;
   lik_expo(D->n, D->x, p, &loglik);
   *sol = -2 * loglik;
}

void nlm_expo(int *n, double *x, double *l, double *dev)
{
/* the function called from R */
   int *itrmcd, *np, term_code, N;
   double *xpls, *fpls, *gpls, est, sol, grad, *a, *wrk;
   TITI *D, data;
   extern void fcn_expo(int, double *, double *, TITI *);

   N = 1;
   np = &N;
   D = &data;
   D->n = n;
   D->x = x;
   itrmcd = &term_code;
   xpls = &est;
   fpls = &sol;
   gpls = &grad;

   a = (double*)malloc(*np * *np * sizeof(double));
   wrk = (double*)malloc(*np * 9 * sizeof(double));
   optif0(*np, *np, l, (fcn_p) fcn_expo, D,
	 xpls, fpls, gpls, itrmcd, a, wrk);
   *l = *xpls;
   *dev = *fpls;
}


### Here is the R function that calls the C code:
nlmexpo <- function(x)
{
     n <- length(x)
     l <- 1
     dev <- 0.1
     c1 <- c2 <- -8
     .C("nlm_expo", as.integer(n), as.double(x),
        as.double(l), as.double(dev),
        NAOK = TRUE, PACKAGE = "apex")
}

I have solved (apparently) my problem: instead of optif0, calling optif9 
works fine. I use something like:

   optif9(*np, *np, l, (fcn_p) fcn_expo, (fcn_p) 0, (d2fcn_p) 0,
	 D, typsiz, 1, 1, 1, &msg, -1 /* = ndigit */, 1000,
	 0 /* = iagflg */, 0, -1.0, 1.e-6, 0.1, 1.e-6, xpls, fpls,
	 gpls, itrmcd, a, wrk, &itncnt);

inspired from what is in nlmefit.c from package nlme (also found in an 
older mail from Adrian Trapletti on R-devel).

It sounds that optif0 is not used anywhere in R.

EP

Emmanuel Paradis wrote:
> Dear All,
> 
> I am trying to use the function optif0 (in main/uncmin.c) from the
> latest R distribution. The reason is that I have a quite complicated
> likelihood function which is coded in C, and I would like to optimize it 
> directly.
> 
> To see how this works, I have tried with a very simple example: 
> optimizing the likelihood of a sample using an exponential distribution. 
> I have tried several solutions but none worked. I paste below the 
> functions that come the closest to what should work. The compilation is 
> fine and the call from R too. It seems that the call to optif0 does not 
> do anything whereas everything else works.