| 1 | |
|---|
| 2 | #include "merger.h" |
|---|
| 3 | #include "../estim/arx.h" |
|---|
| 4 | |
|---|
| 5 | namespace bdm |
|---|
| 6 | { |
|---|
| 7 | vec merger_base::merge_points ( mat &lW ) { |
|---|
| 8 | int nu=lW.rows(); |
|---|
| 9 | vec result; |
|---|
| 10 | ivec indW; |
|---|
| 11 | bool infexist; |
|---|
| 12 | switch ( METHOD ) { |
|---|
| 13 | case ARITHMETIC: |
|---|
| 14 | result= log ( sum ( exp ( lW ) ) ); //ugly! |
|---|
| 15 | break; |
|---|
| 16 | case GEOMETRIC: |
|---|
| 17 | result= sum ( lW ) /nu; |
|---|
| 18 | break; |
|---|
| 19 | case LOGNORMAL: |
|---|
| 20 | vec sumlW=sum ( lW ) ; |
|---|
| 21 | indW=find((sumlW<inf) & (sumlW>-inf)); |
|---|
| 22 | infexist=(indW.size()<lW.cols()); |
|---|
| 23 | vec mu; |
|---|
| 24 | vec lam; |
|---|
| 25 | if (infexist){ |
|---|
| 26 | mu = sumlW(indW) /nu; //mean of logs |
|---|
| 27 | // |
|---|
| 28 | mat validlW=lW.get_cols(indW); |
|---|
| 29 | lam = sum ( pow ( validlW-outer_product ( ones ( validlW.rows() ),mu ),2 ) ); |
|---|
| 30 | } |
|---|
| 31 | else { |
|---|
| 32 | mu = sum ( lW ) /nu; //mean of logs |
|---|
| 33 | lam = sum ( pow ( lW-outer_product ( ones ( lW.rows() ),mu ),2 ) ); |
|---|
| 34 | } |
|---|
| 35 | // |
|---|
| 36 | double coef=0.0; |
|---|
| 37 | vec sq2bl=sqrt ( 2*beta*lam ); //this term is everywhere |
|---|
| 38 | switch ( nu ) { |
|---|
| 39 | case 2: |
|---|
| 40 | coef= ( 1-0.5*sqrt ( ( 4.0*beta-3.0 ) /beta ) ); |
|---|
| 41 | result =coef*sq2bl + mu ; |
|---|
| 42 | break; |
|---|
| 43 | // case 4: == can be done similar to case 2 - is it worth it??? |
|---|
| 44 | default: // see accompanying document merge_lognorm_derivation.lyx |
|---|
| 45 | coef = sqrt(1-(nu+1)/(2*beta*nu)); |
|---|
| 46 | result =log(besselk((nu-3)/2, sq2bl*coef))-log(besselk((nu-3)/2, sq2bl)) + mu; |
|---|
| 47 | break; |
|---|
| 48 | } |
|---|
| 49 | break; |
|---|
| 50 | } |
|---|
| 51 | if (infexist){ |
|---|
| 52 | vec tmp =-inf*ones(lW.cols()); |
|---|
| 53 | set_subvector(tmp, indW, result); |
|---|
| 54 | return tmp; |
|---|
| 55 | } |
|---|
| 56 | else {return result;} |
|---|
| 57 | } |
|---|
| 58 | |
|---|
| 59 | void merger_mix::merge ( ) |
|---|
| 60 | { |
|---|
| 61 | Array<vec> &Smp = eSmp._samples(); //aux |
|---|
| 62 | vec &w = eSmp._w(); //aux |
|---|
| 63 | |
|---|
| 64 | mat Smp_ex =ones ( dim +1,Npoints ); // Extended samples for the ARX model - the last row is ones |
|---|
| 65 | for ( int i=0;i<Npoints;i++ ) { set_col_part ( Smp_ex,i,Smp ( i ) );} |
|---|
| 66 | |
|---|
| 67 | if ( DBG ) *dbg_file << Name ( "Smp_0" ) << Smp_ex; |
|---|
| 68 | |
|---|
| 69 | // Stuff for merging |
|---|
| 70 | vec lw_src ( Npoints ); // weights of the ith source |
|---|
| 71 | vec lw_mix ( Npoints ); // weights of the approximating mixture |
|---|
| 72 | vec lw ( Npoints ); // tmp |
|---|
| 73 | mat lW=zeros ( Nsources,Npoints ); // array of weights of all sources |
|---|
| 74 | vec vec0 ( 0 ); |
|---|
| 75 | |
|---|
| 76 | //initialize importance weights |
|---|
| 77 | lw_mix = 1.0; // assuming uniform grid density -- otherwise |
|---|
| 78 | |
|---|
| 79 | // Initial component in the mixture model |
|---|
| 80 | mat V0=1e-8*eye ( dim +1 ); |
|---|
| 81 | ARX A0; A0.set_statistics ( dim, V0 ); //initial guess of Mix: |
|---|
| 82 | |
|---|
| 83 | Mix.init ( &A0, Smp_ex, Ncoms ); |
|---|
| 84 | //Preserve initial mixture for repetitive estimation via flattening |
|---|
| 85 | MixEF Mix_init ( Mix ); |
|---|
| 86 | |
|---|
| 87 | // ============= MAIN LOOP ================== |
|---|
| 88 | bool converged=false; |
|---|
| 89 | int niter = 0; |
|---|
| 90 | char dbg_str[100]; |
|---|
| 91 | |
|---|
| 92 | emix* Mpred=Mix.epredictor ( ); |
|---|
| 93 | vec Mix_pdf ( Npoints ); |
|---|
| 94 | while ( !converged ) |
|---|
| 95 | { |
|---|
| 96 | //Re-estimate Mix |
|---|
| 97 | //Re-Initialize Mixture model |
|---|
| 98 | Mix.flatten ( &Mix_init ); |
|---|
| 99 | Mix.bayesB ( Smp_ex, w*Npoints ); |
|---|
| 100 | delete Mpred; |
|---|
| 101 | Mpred = Mix.epredictor ( ); // Allocation => must be deleted at the end!! |
|---|
| 102 | Mpred->set_rv ( rv ); //the predictor predicts rv of this merger |
|---|
| 103 | |
|---|
| 104 | // This will be active only later in iterations!!! |
|---|
| 105 | if ( 1./sum_sqr ( w ) <effss_coef*Npoints ) |
|---|
| 106 | { |
|---|
| 107 | // Generate new samples |
|---|
| 108 | eSmp.set_samples ( Mpred ); |
|---|
| 109 | for ( int i=0;i<Npoints;i++ ) |
|---|
| 110 | { |
|---|
| 111 | //////////// !!!!!!!!!!!!! |
|---|
| 112 | //if ( Smp ( i ) ( 2 ) <0 ) {Smp ( i ) ( 2 ) = 0.01; } |
|---|
| 113 | set_col_part ( Smp_ex,i,Smp ( i ) ); |
|---|
| 114 | //Importance of the mixture |
|---|
| 115 | //lw_mix ( i ) =Mix.logpred (Smp_ex.get_col(i) ); |
|---|
| 116 | lw_mix ( i ) = Mpred->evallog ( Smp ( i ) ); |
|---|
| 117 | } |
|---|
| 118 | if ( DBG ) |
|---|
| 119 | { |
|---|
| 120 | cout<<"Resampling =" << 1./sum_sqr ( w ) << endl; |
|---|
| 121 | cout << Mix._e()->mean() <<endl; |
|---|
| 122 | cout << sum ( Smp_ex,2 ) /Npoints <<endl; |
|---|
| 123 | cout << Smp_ex*Smp_ex.T() /Npoints << endl; |
|---|
| 124 | } |
|---|
| 125 | } |
|---|
| 126 | if ( DBG ) |
|---|
| 127 | { |
|---|
| 128 | sprintf ( dbg_str,"Mpred_mean%d",niter ); |
|---|
| 129 | *dbg_file << Name ( dbg_str ) << Mpred->mean(); |
|---|
| 130 | sprintf ( dbg_str,"Mpred_var%d",niter ); |
|---|
| 131 | *dbg_file << Name ( dbg_str ) << Mpred->variance(); |
|---|
| 132 | |
|---|
| 133 | |
|---|
| 134 | sprintf ( dbg_str,"Mpdf%d",niter ); |
|---|
| 135 | for ( int i=0;i<Npoints;i++ ) {Mix_pdf ( i ) = Mix.logpred ( Smp_ex.get_col ( i ) );} |
|---|
| 136 | *dbg_file << Name ( dbg_str ) << Mix_pdf; |
|---|
| 137 | |
|---|
| 138 | sprintf ( dbg_str,"Smp%d",niter ); |
|---|
| 139 | *dbg_file << Name ( dbg_str ) << Smp_ex; |
|---|
| 140 | |
|---|
| 141 | } |
|---|
| 142 | //Importace weighting |
|---|
| 143 | for ( int i=0;i<mpdfs.length();i++ ) |
|---|
| 144 | { |
|---|
| 145 | lw_src=0.0; |
|---|
| 146 | //======== Same RVs =========== |
|---|
| 147 | //Split according to dependency in rvs |
|---|
| 148 | if ( mpdfs ( i )->dimension() ==dim ) |
|---|
| 149 | { |
|---|
| 150 | // no need for conditioning or marginalization |
|---|
| 151 | lw_src = mpdfs ( i )->_epdf().evallog_m ( Smp ); |
|---|
| 152 | } |
|---|
| 153 | else |
|---|
| 154 | { |
|---|
| 155 | // compute likelihood of marginal on the conditional variable |
|---|
| 156 | if ( mpdfs ( i )->dimensionc() >0 ) |
|---|
| 157 | { |
|---|
| 158 | // Make marginal on rvc_i |
|---|
| 159 | epdf* tmp_marg = Mpred->marginal ( mpdfs ( i )->_rvc() ); |
|---|
| 160 | //compute vector of lw_src |
|---|
| 161 | for ( int k=0;k<Npoints;k++ ) |
|---|
| 162 | { |
|---|
| 163 | // Here val of tmp_marg = cond of mpdfs(i) ==> calling dls->get_cond |
|---|
| 164 | lw_src ( k ) += tmp_marg->evallog ( dls ( i )->get_cond ( Smp ( k ) ) ); |
|---|
| 165 | } |
|---|
| 166 | delete tmp_marg; |
|---|
| 167 | |
|---|
| 168 | // sprintf ( str,"marg%d",niter ); |
|---|
| 169 | // *dbg << Name ( str ) << lw_src; |
|---|
| 170 | |
|---|
| 171 | } |
|---|
| 172 | // Compute likelihood of the missing variable |
|---|
| 173 | if ( dim > ( mpdfs ( i )->dimension() + mpdfs ( i )->dimensionc() ) ) |
|---|
| 174 | { |
|---|
| 175 | /////////////// |
|---|
| 176 | // There are variales unknown to mpdfs(i) : rvzs |
|---|
| 177 | mpdf* tmp_cond = Mpred->condition ( rvzs ( i ) ); |
|---|
| 178 | // Compute likelihood |
|---|
| 179 | vec lw_dbg=lw_src; |
|---|
| 180 | for ( int k= 0; k<Npoints; k++ ) |
|---|
| 181 | { |
|---|
| 182 | lw_src ( k ) += log ( |
|---|
| 183 | tmp_cond->evallogcond ( |
|---|
| 184 | zdls ( i )->pushdown ( Smp ( k ) ), |
|---|
| 185 | zdls ( i )->get_cond ( Smp ( k ) ) ) ); |
|---|
| 186 | if ( !std::isfinite ( lw_src ( k ) ) ) |
|---|
| 187 | { |
|---|
| 188 | lw_src ( k ) = -1e16; cout << "!"; |
|---|
| 189 | } |
|---|
| 190 | } |
|---|
| 191 | delete tmp_cond; |
|---|
| 192 | } |
|---|
| 193 | // Compute likelihood of the partial source |
|---|
| 194 | for ( int k= 0; k<Npoints; k++ ) |
|---|
| 195 | { |
|---|
| 196 | mpdfs ( i )->condition ( dls ( i )->get_cond ( Smp ( k ) ) ); |
|---|
| 197 | lw_src ( k ) += mpdfs ( i )->_epdf().evallog ( dls ( i )->pushdown ( Smp ( k ) ) ); |
|---|
| 198 | } |
|---|
| 199 | |
|---|
| 200 | } |
|---|
| 201 | // it_assert_debug(std::isfinite(sum(lw_src)),"bad"); |
|---|
| 202 | lW.set_row ( i, lw_src ); // do not divide by mix |
|---|
| 203 | } |
|---|
| 204 | lw = merger_base::merge_points ( lW ); //merge |
|---|
| 205 | |
|---|
| 206 | //Importance weighting |
|---|
| 207 | lw -= lw_mix; // hoping that it is not numerically sensitive... |
|---|
| 208 | w = exp ( lw-max ( lw ) ); |
|---|
| 209 | |
|---|
| 210 | //renormalize |
|---|
| 211 | double sumw =sum ( w ); |
|---|
| 212 | if ( std::isfinite ( sumw ) ) |
|---|
| 213 | { |
|---|
| 214 | w = w/sumw; |
|---|
| 215 | } |
|---|
| 216 | else |
|---|
| 217 | { |
|---|
| 218 | it_file itf ( "merg_err.it" ); |
|---|
| 219 | itf << Name ( "w" ) << w; |
|---|
| 220 | } |
|---|
| 221 | |
|---|
| 222 | if ( DBG ) |
|---|
| 223 | { |
|---|
| 224 | sprintf ( dbg_str,"lW%d",niter ); |
|---|
| 225 | *dbg_file << Name ( dbg_str ) << lW; |
|---|
| 226 | sprintf ( dbg_str,"w%d",niter ); |
|---|
| 227 | *dbg_file << Name ( dbg_str ) << w; |
|---|
| 228 | sprintf ( dbg_str,"lw_m%d",niter ); |
|---|
| 229 | *dbg_file << Name ( dbg_str ) << lw_mix; |
|---|
| 230 | } |
|---|
| 231 | // ==== stopping rule === |
|---|
| 232 | niter++; |
|---|
| 233 | converged = ( niter>stop_niter ); |
|---|
| 234 | } |
|---|
| 235 | delete Mpred; |
|---|
| 236 | // cout << endl; |
|---|
| 237 | |
|---|
| 238 | } |
|---|
| 239 | |
|---|
| 240 | // DEFAULTS FOR MERGER_BASE |
|---|
| 241 | const MERGER_METHOD merger_base::DFLT_METHOD=LOGNORMAL; |
|---|
| 242 | const double merger_base::DFLT_beta=1.2; |
|---|
| 243 | // DEFAULTS FOR MERGER_MIX |
|---|
| 244 | const int merger_mix::DFLT_Ncoms=10; |
|---|
| 245 | const double merger_mix::DFLT_effss_coef=0.5; |
|---|
| 246 | |
|---|
| 247 | } |
|---|
