ADMB Documentation  -a65f1c97
 All Classes Namespaces Files Functions Variables Typedefs Enumerations Enumerator Friends Macros Groups Pages
df1b2im5.cpp
Go to the documentation of this file.
1 /*
2  * $Id$
3  *
4  * Author: David Fournier
5  * Copyright (c) 2008-2012 Regents of the University of California
6  */
11 #include <admodel.h>
12 #include <df1b2fun.h>
13 #include <adrndeff.h>
14 #ifndef OPT_LIB
15  #include <cassert>
16  #include <climits>
17 #endif
18 
24  (const dvector& x,const dvector& u0,const dmatrix& Hess,
25  const dvector& _xadjoint,const dvector& _uadjoint,
26  const dmatrix& _Hessadjoint,function_minimizer * pmin)
27 {
28  ADUNCONST(dvector,xadjoint)
29  ADUNCONST(dvector,uadjoint)
30  //ADUNCONST(dmatrix,Hessadjoint)
31 #if !defined(OPT_LIB) && (__cplusplus >= 201103L)
32  const int xs = [](unsigned int size)->int
33  {
34  assert(size <= INT_MAX);
35  return static_cast<int>(size);
36  }(x.size());
37  const int us = [](unsigned int size)->int
38  {
39  assert(size <= INT_MAX);
40  return static_cast<int>(size);
41  }(u0.size());
42 #else
43  const int xs = static_cast<int>(x.size());
44  const int us = static_cast<int>(u0.size());
45 #endif
47  int nsc=pmin->lapprox->num_separable_calls;
48  const ivector lrea = (*pmin->lapprox->num_local_re_array)(1,nsc);
49  int hroom = sum(square(lrea));
50  int nvar = xs + us + hroom;
51  independent_variables y(1,nvar);
52 
53  // need to set random effects active together with whatever
54  // init parameters should be active in this phase
57  /*int onvar=*/initial_params::nvarcalc();
58  initial_params::xinit(y); // get the initial values into the
59  // do we need this next line?
60  y(1,xs)=x;
61 
62  // contribution for quadratic prior
64  {
65  //Hess+=quadratic_prior::get_cHessian_contribution();
66  int & vxs = (int&)(xs);
68  }
69  // Here need hooks for sparse matrix structures
70 
71  dvar3_array & block_diagonal_vhessian=
73  block_diagonal_vhessian.initialize();
74  dvar3_array& block_diagonal_ch=
76  //dvar3_array(*pmin->lapprox->block_diagonal_ch);
77  int ii=xs+us+1;
79  for (int ic=1;ic<=nsc;ic++)
80  {
81  int lus=lrea(ic);
82  for (int i=1;i<=lus;i++)
83  for (int j=1;j<=lus;j++)
84  y(ii++)=bdH(ic)(i,j);
85  }
86 
87  dvector g(1,nvar);
88  gradcalc(0,g);
91  //initial_params::stddev_vscale(d,vy);
92  ii=xs+us+1;
94  {
95  cerr << "can't do importance sampling with bounded random effects"
96  " at present" << endl;
97  ad_exit(1);
98  }
99  else
100  {
101  for (int ic=1;ic<=nsc;ic++)
102  {
103  int lus=lrea(ic);
104  for (int i=1;i<=lus;i++)
105  {
106  for (int j=1;j<=lus;j++)
107  {
108  block_diagonal_vhessian(ic,i,j)=vy(ii++);
109  }
110  }
111  block_diagonal_ch(ic)=
112  choleski_decomp(inv(block_diagonal_vhessian(ic)));
113  }
114  }
115 
116  int nsamp=pmin->lapprox->num_importance_samples;
117 
118  dvariable vf=0.0;
119 
120  dvar_vector sample_value(1,nsamp);
121  sample_value.initialize();
122 
123  dvar_vector tau(1,us);;
124  int is;
125  for (is=1;is<=nsamp;is++)
126  {
127  int offset=0;
129  for (int ic=1;ic<=nsc;ic++)
130  {
131  int lus=lrea(ic);
132  tau(offset+1,offset+lus).shift(1)=block_diagonal_ch(ic)*
133  (*pmin->lapprox->antiepsilon)(is);
134  offset+=lus;
135  }
136 
137  // have to reorder the terms to match the block diagonal hessian
138  imatrix & ls=*(pmin->lapprox->block_diagonal_re_list);
139  int mmin=ls.indexmin();
140  int mmax=ls.indexmax();
141 
142  ii=1;
143  for (int i=mmin;i<=mmax;i++)
144  {
145  int cmin=ls(i).indexmin();
146  int cmax=ls(i).indexmax();
147  for (int j=cmin;j<=cmax;j++)
148  {
149  vy(ls(i,j))+=tau(ii++);
150  }
151  }
152  if (ii-1 != us)
153  {
154  cerr << "error in interface" << endl;
155  ad_exit(1);
156  }
157  initial_params::reset(vy); // get the values into the model
158  ii=1;
159  for (int i=mmin;i<=mmax;i++)
160  {
161  int cmin=ls(i).indexmin();
162  int cmax=ls(i).indexmax();
163  for (int j=cmin;j<=cmax;j++)
164  {
165  vy(ls(i,j))-=tau(ii++);
166  }
167  }
168 
170  pmin->AD_uf_outer();
171 
172  if (pmin->lapprox->use_outliers==0)
173  {
174  // assumes that all separable calls have the same number
175  // of random effects
176  double neps=0.5*nsc*norm2((*pmin->lapprox->antiepsilon)(is));
177 
178  (*pmin->lapprox->importance_sampling_values)(is)=
180 
181  (*pmin->lapprox->importance_sampling_weights)(is)=neps;
182 
183  sample_value(is)=*objective_function_value::pobjfun
184  -neps;
185  }
186  else
187  {
188  dvector& e=pmin->lapprox->epsilon(is);
189  double neps=-sum(log(.95*exp(-0.5*square(e))+
190  0.0166666667*exp(-square(e)/18.0)));
191 
192  (*pmin->lapprox->importance_sampling_values)(is)=
194 
195  sample_value(is)=*objective_function_value::pobjfun
196  -neps;
197  }
198  }
199 
200  nsc=pmin->lapprox->num_separable_calls;
201  dmatrix weights(1,nsc,1,nsamp);
202  for (is=1;is<=nsamp;is++)
203  {
204  int offset=0;
205  for (int ic=1;ic<=nsc;ic++)
206  {
207  int lus=lrea(ic);
208  // assumes that all spearable calls have the same number of
209  // random effects
210  dvector e= (*pmin->lapprox->antiepsilon)(is);
211  offset+=lus;
212  if (pmin->lapprox->use_outliers==0)
213  {
214  weights(ic,is)=0.5*norm2(e);
215  }
216  else
217  {
218  weights(ic,is)=-sum(log(.95*exp(-0.5*square(e))+
219  0.0166666667*exp(-square(e)/18.0)));
220  }
221  }
222  }
223  dvar_vector lcomp(1,nsc);
224  for (int isc=1;isc<=nsc;isc++)
225  {
226  dvar_vector & comps=
228 
229  dvar_vector diff=comps-weights(isc);
230  double dmin=min(value(diff));
231  lcomp(isc)=dmin-log(mean(exp(dmin-diff)));
232  }
233 
234  //double ns=lcomp.indexmax()-lcomp.indexmin()+1;
235  //double min_vf=min(value(lcomp));
236  vf= sum(lcomp);
237  vf-=us*0.91893853320467241;
238 
239  int sgn=0;
240  dvariable ld=0.0;
242  {
243  for (int ic=1;ic<=nsc;ic++)
244  {
245  ld+=ln_det(block_diagonal_vhessian(ic),sgn);
246  }
247  ld*=0.5;
248  }
249  else
250  {
251  for (int ic=1;ic<=nsc;ic++)
252  {
253  ld+=ln_det_choleski(block_diagonal_vhessian(ic));
254  }
255  ld*=0.5;
256  }
257 
258  vf+=ld;
259 
260  double f=value(vf);
261  gradcalc(nvar,g);
262 
263  // put uhat back into the model
265  vy(xs+1,xs+us).shift(1)=u0;
266  initial_params::reset(vy); // get the values into the model
268 
269  ii=1;
270  for (int i=1;i<=xs;i++)
271  xadjoint(i)=g(ii++);
272  for (int i=1;i<=us;i++)
273  uadjoint(i)=g(ii++);
274  for (int ic=1;ic<=nsc;ic++)
275  {
276  int lus=lrea(ic);
277  for (int i=1;i<=lus;i++)
278  {
279  for (int j=1;j<=lus;j++)
280  {
281  (*pmin->lapprox->block_diagonal_vhessianadjoint)(ic)(i,j)=g(ii++);
282  }
283  }
284  }
285  return f;
286 }
laplace_approximation_calculator * lapprox
Definition: admodel.h:1862
int indexmax() const
Definition: imatrix.h:142
Description not yet available.
Definition: imatrix.h:69
static void set_NO_DERIVATIVES(void)
Disable accumulation of derivative information.
Definition: gradstrc.cpp:641
Description not yet available.
dvar_vector & shift(int min)
Description not yet available.
Definition: fvar_arr.cpp:28
static void set_active_random_effects(void)
Definition: model.cpp:267
int indexmin() const
Definition: imatrix.h:138
#define x
#define ADUNCONST(type, obj)
Creates a shallow copy of obj that is not CONST.
Definition: fvar.hpp:140
Vector of double precision numbers.
Definition: dvector.h:50
dvar_matrix * importance_sampling_components
Definition: adrndeff.h:224
void initialize(void)
Description not yet available.
Definition: f3arr.cpp:17
double sum(const d3_array &darray)
Author: David Fournier Copyright (c) 2008-2012 Regents of the University of California.
Definition: d3arr.cpp:21
exitptr ad_exit
Definition: gradstrc.cpp:53
#define dmin(a, b)
Definition: cbivnorm.cpp:190
static dvariable reset(const dvar_vector &x)
Definition: model.cpp:345
ADMB variable vector.
Definition: fvar.hpp:2172
double mean(const dvector &vec)
Returns computed mean of vec.
Definition: cranfill.cpp:43
void gradcalc(int nvar, const dvector &g)
Definition: sgradclc.cpp:77
df1_one_matrix choleski_decomp(const df1_one_matrix &MM)
Definition: df11fun.cpp:606
static int nvarcalc()
Definition: model.cpp:152
ivector sgn(const dvector &v)
Author: David Fournier Copyright (c) 2008-2012 Regents of the University of California.
Definition: dvect24.cpp:11
static int no_ln_det_choleski_flag
Definition: fvar.hpp:8841
dvar3_array * block_diagonal_vch
Definition: adrndeff.h:228
prnstream & endl(prnstream &)
Description not yet available.
Definition: fvar.hpp:1937
Array of integers(int) with indexes from index_min to indexmax.
Definition: ivector.h:50
#define min(a, b)
Definition: cbivnorm.cpp:188
double calculate_importance_sample_block_diagonal_option_antithetical(const dvector &x, const dvector &u0, const dmatrix &Hess, const dvector &_xadjoint, const dvector &_uadjoint, const dmatrix &_Hessadjoint, function_minimizer *pmin)
Description not yet available.
Definition: df1b2im5.cpp:24
static objective_function_value * pobjfun
Definition: admodel.h:2394
Description not yet available.
static void xinit(const dvector &x)
Definition: model.cpp:226
d3_array exp(const d3_array &arr3)
Returns d3_array results with computed exp from elements in arr3.
Definition: d3arr2a.cpp:28
Description not yet available.
Definition: fvar.hpp:2819
Description not yet available.
Definition: fvar.hpp:4197
double norm2(const d3_array &a)
Return sum of squared elements in a.
Definition: d3arr2a.cpp:167
d3_array * block_diagonal_vhessianadjoint
Definition: adrndeff.h:232
double ln_det(const dmatrix &m1, int &sgn)
Compute log determinant of a constant matrix.
Definition: dmat3.cpp:536
unsigned int size() const
Get number of elements in array.
Definition: dvector.h:209
dvar3_array * block_diagonal_vhessian
Definition: adrndeff.h:233
static int have_bounded_random_effects
Definition: df1b2fun.h:1353
virtual void AD_uf_outer()
Definition: xmodelm4.cpp:39
double ln_det_choleski(const banded_symmetric_dmatrix &MM, int &ierr)
Definition: dmat38.cpp:218
Description not yet available.
Description not yet available.
Definition: admodel.h:1850
static void set_YES_DERIVATIVES(void)
Enable accumulation of derivative information.
Definition: gradstrc.cpp:650
static void get_cHessian_contribution(dmatrix, int)
Description not yet available.
Definition: quadpri.cpp:591
void initialize(const dvector &ww)
Description not yet available.
Definition: fvar_a24.cpp:63
dvector value(const df1_one_vector &v)
Definition: df11fun.cpp:69
Description not yet available.
Definition: fvar.hpp:3727
static int get_num_quadratic_prior(void)
Definition: df1b2fun.h:1916
double square(const double value)
Return square of value; constant object.
Definition: d3arr4.cpp:16
Fundamental data type for reverse mode automatic differentiation.
Definition: fvar.hpp:1518
df1_one_variable inv(const df1_one_variable &x)
Definition: df11fun.cpp:384
d3_array log(const d3_array &arr3)
Author: David Fournier Copyright (c) 2008-2012 Regents of the University of California.
Definition: d3arr2a.cpp:13
static void set_inactive_only_random_effects(void)
Definition: model.cpp:259