ADMB Documentation  11.2.2826
 All Classes Files Functions Variables Typedefs Friends Defines
df1b2im4.cpp
Go to the documentation of this file.
00001 /*
00002  * $Id: df1b2im4.cpp 2684 2014-11-18 19:50:21Z johnoel $
00003  *
00004  * Author: David Fournier
00005  * Copyright (c) 2008-2012 Regents of the University of California
00006  */
00011 #  include <admodel.h>
00012 #  include <df1b2fun.h>
00013 #  include <adrndeff.h>
00014 
00019 double calculate_importance_sample_block_diagonal_option2(const dvector& x,
00020   const dvector& u0,const dmatrix& Hess,const dvector& _xadjoint,
00021   const dvector& _uadjoint,const dmatrix& _Hessadjoint,
00022   function_minimizer * pmin)
00023 {
00024   ADUNCONST(dvector,xadjoint)
00025   ADUNCONST(dvector,uadjoint)
00026   //ADUNCONST(dmatrix,Hessadjoint)
00027   const int xs=x.size();
00028   const int us=u0.size();
00029   gradient_structure::set_NO_DERIVATIVES();
00030   int nsc=pmin->lapprox->num_separable_calls;
00031   const ivector lrea = (*pmin->lapprox->num_local_re_array)(1,nsc);
00032   int hroom =  int(sum(square(lrea)));
00033   int nvar=x.size()+u0.size()+hroom;
00034   independent_variables y(1,nvar);
00035 
00036   // need to set random effects active together with whatever
00037   // init parameters should be active in this phase
00038   initial_params::set_inactive_only_random_effects();
00039   initial_params::set_active_random_effects();
00040   /*int onvar=*/initial_params::nvarcalc();
00041   initial_params::xinit(y);    // get the initial values into the
00042   // do we need this next line?
00043   y(1,xs)=x;
00044 
00045   // contribution for quadratic prior
00046   if (quadratic_prior::get_num_quadratic_prior()>0)
00047   {
00048     //Hess+=quadratic_prior::get_cHessian_contribution();
00049     int & vxs = (int&)(xs);
00050     quadratic_prior::get_cHessian_contribution(Hess,vxs);
00051   }
00052  // Here need hooks for sparse matrix structures
00053 
00054   dvar3_array & block_diagonal_vhessian=
00055     *pmin->lapprox->block_diagonal_vhessian;
00056   block_diagonal_vhessian.initialize();
00057   dvar3_array& block_diagonal_ch=
00058     *pmin->lapprox->block_diagonal_vch;
00059     //dvar3_array(*pmin->lapprox->block_diagonal_ch);
00060   int ii=xs+us+1;
00061   d3_array& bdH=(*pmin->lapprox->block_diagonal_hessian);
00062   for (int ic=1;ic<=nsc;ic++)
00063   {
00064     int lus=lrea(ic);
00065     for (int i=1;i<=lus;i++)
00066       for (int j=1;j<=lus;j++)
00067         y(ii++)=bdH(ic)(i,j);
00068   }
00069 
00070   dvector g(1,nvar);
00071   gradcalc(0,g);
00072   gradient_structure::set_YES_DERIVATIVES();
00073   dvar_vector vy=dvar_vector(y);
00074   //initial_params::stddev_vscale(d,vy);
00075   ii=xs+us+1;
00076   if (initial_df1b2params::have_bounded_random_effects)
00077   {
00078     cerr << "can't do importance sampling with bounded random effects"
00079      " at present" << endl;
00080     ad_exit(1);
00081   }
00082   else
00083   {
00084     for (int ic=1;ic<=nsc;ic++)
00085     {
00086       int lus=lrea(ic);
00087       for (int i=1;i<=lus;i++)
00088       {
00089         for (int j=1;j<=lus;j++)
00090         {
00091           block_diagonal_vhessian(ic,i,j)=vy(ii++);
00092         }
00093       }
00094       if (lus>0)
00095         block_diagonal_ch(ic)=
00096           choleski_decomp(inv(block_diagonal_vhessian(ic)));
00097     }
00098   }
00099 
00100    int nsamp=pmin->lapprox->num_importance_samples;
00101 
00102    dvariable vf=0.0;
00103 
00104    dvar_vector sample_value(1,nsamp);
00105    sample_value.initialize();
00106 
00107    dvar_vector tau(1,us);;
00108    int is;
00109    for (is=1;is<=nsamp;is++)
00110    {
00111      int offset=0;
00112      pmin->lapprox->importance_sampling_counter=is;
00113      for (int ic=1;ic<=nsc;ic++)
00114      {
00115        int lus=lrea(ic);
00116        if (lus>0)
00117          tau(offset+1,offset+lus).shift(1)=block_diagonal_ch(ic)*
00118            pmin->lapprox->epsilon(is)(offset+1,offset+lus).shift(1);
00119        offset+=lus;
00120      }
00121 
00122      // have to reorder the terms to match the block diagonal hessian
00123      imatrix & ls=*(pmin->lapprox->block_diagonal_re_list);
00124      int mmin=ls.indexmin();
00125      int mmax=ls.indexmax();
00126 
00127      ii=1;
00128      for (int i=mmin;i<=mmax;i++)
00129      {
00130        int cmin=ls(i).indexmin();
00131        int cmax=ls(i).indexmax();
00132        for (int j=cmin;j<=cmax;j++)
00133        {
00134          vy(ls(i,j))+=tau(ii++);
00135        }
00136      }
00137      if (ii-1 != us)
00138      {
00139        cerr << "error in interface" << endl;
00140        ad_exit(1);
00141      }
00142      initial_params::reset(vy);    // get the values into the model
00143      ii=1;
00144      for (int i=mmin;i<=mmax;i++)
00145      {
00146        int cmin=ls(i).indexmin();
00147        int cmax=ls(i).indexmax();
00148        for (int j=cmin;j<=cmax;j++)
00149        {
00150          vy(ls(i,j))-=tau(ii++);
00151        }
00152      }
00153 
00154      *objective_function_value::pobjfun=0.0;
00155      pmin->AD_uf_outer();
00156 
00157      if (pmin->lapprox->use_outliers==0)
00158      {
00159        double neps=0.5*norm2(pmin->lapprox->epsilon(is));
00160 
00161        (*pmin->lapprox->importance_sampling_values)(is)=
00162            neps-value(*objective_function_value::pobjfun);
00163 
00164        (*pmin->lapprox->importance_sampling_weights)(is)=neps;
00165 
00166         sample_value(is)=*objective_function_value::pobjfun
00167           -neps;
00168      }
00169      else
00170      {
00171        dvector& e=pmin->lapprox->epsilon(is);
00172        double neps=-sum(log(.95*exp(-0.5*square(e))+
00173           0.0166666667*exp(-square(e)/18.0)));
00174 
00175        (*pmin->lapprox->importance_sampling_values)(is)=
00176          neps-value(*objective_function_value::pobjfun);
00177 
00178        sample_value(is)=*objective_function_value::pobjfun
00179          -neps;
00180      }
00181    }
00182 
00183    nsc=pmin->lapprox->num_separable_calls;
00184    dmatrix weights(1,nsc,1,nsamp);
00185    for (is=1;is<=nsamp;is++)
00186    {
00187      int offset=0;
00188      for (int ic=1;ic<=nsc;ic++)
00189      {
00190        int lus=lrea(ic);
00191        dvector e= pmin->lapprox->epsilon(is)(offset+1,offset+lus).shift(1);
00192        offset+=lus;
00193        if (pmin->lapprox->use_outliers==0)
00194        {
00195          weights(ic,is)=0.5*norm2(e);
00196        }
00197        else
00198        {
00199          weights(ic,is)=-sum(log(.95*exp(-0.5*square(e))+
00200           0.0166666667*exp(-square(e)/18.0)));
00201        }
00202      }
00203    }
00204    dvar_vector lcomp(1,nsc);
00205    for (int isc=1;isc<=nsc;isc++)
00206    {
00207      dvar_vector & comps=
00208        (*pmin->lapprox->importance_sampling_components)(isc);
00209 
00210      dvar_vector diff=comps-weights(isc);
00211      double dmin=min(value(diff));
00212      lcomp(isc)=dmin-log(mean(exp(dmin-diff)));
00213    }
00214 
00215    //double ns=lcomp.indexmax()-lcomp.indexmin()+1;
00216    //double min_vf=min(value(lcomp));
00217    vf= sum(lcomp);
00218    vf-=us*0.91893853320467241;
00219 
00220    int sgn=0;
00221    dvariable ld=0.0;
00222    if (ad_comm::no_ln_det_choleski_flag)
00223    {
00224      for (int ic=1;ic<=nsc;ic++)
00225      {
00226        ld+=ln_det(block_diagonal_vhessian(ic),sgn);
00227      }
00228      ld*=0.5;
00229    }
00230    else
00231    {
00232      for (int ic=1;ic<=nsc;ic++)
00233      {
00234        if (allocated(block_diagonal_vhessian(ic)))
00235          ld+=ln_det_choleski(block_diagonal_vhessian(ic));
00236      }
00237      ld*=0.5;
00238    }
00239 
00240    vf+=ld;
00241 
00242    double f=value(vf);
00243    gradcalc(nvar,g);
00244 
00245    // put uhat back into the model
00246    gradient_structure::set_NO_DERIVATIVES();
00247    vy(xs+1,xs+us).shift(1)=u0;
00248    initial_params::reset(vy);    // get the values into the model
00249    gradient_structure::set_YES_DERIVATIVES();
00250 
00251   ii=1;
00252   for (int i=1;i<=xs;i++)
00253     xadjoint(i)=g(ii++);
00254   for (int i=1;i<=us;i++)
00255     uadjoint(i)=g(ii++);
00256   for (int ic=1;ic<=nsc;ic++)
00257   {
00258     int lus=lrea(ic);
00259     for (int i=1;i<=lus;i++)
00260     {
00261       for (int j=1;j<=lus;j++)
00262       {
00263         (*pmin->lapprox->block_diagonal_vhessianadjoint)(ic)(i,j)=g(ii++);
00264       }
00265     }
00266   }
00267   return f;
00268 }