| 1 | #include <itpp/itbase.h> |
|---|
| 2 | #include "merger.h" |
|---|
| 3 | #include "arx.h" |
|---|
| 4 | |
|---|
| 5 | vec merger::lognorm_merge ( mat &lW ) { |
|---|
| 6 | int nu=lW.rows(); |
|---|
| 7 | vec mu = sum ( lW ) /nu; //mean of logs |
|---|
| 8 | vec lam = sum ( pow ( lW,2 ) )-nu*pow ( mu,2 ); |
|---|
| 9 | double coef=0.0; |
|---|
| 10 | vec sq2bl=sqrt ( 2*beta*lam ); //this term is everywhere |
|---|
| 11 | switch ( nu ) { |
|---|
| 12 | case 2: |
|---|
| 13 | coef= ( 1-0.5*sqrt ( ( 4*beta-3 ) /beta ) ); |
|---|
| 14 | return exp ( coef*sq2bl + mu ); |
|---|
| 15 | break; |
|---|
| 16 | case 3://Ratio of Bessel |
|---|
| 17 | coef = sqrt ( ( 3*beta-2 ) /3*beta ); |
|---|
| 18 | return elem_mult ( |
|---|
| 19 | elem_div ( besselk ( 0,sq2bl*coef ), besselk ( 0,sq2bl ) ), |
|---|
| 20 | exp ( mu ) ); |
|---|
| 21 | break; |
|---|
| 22 | case 4: |
|---|
| 23 | break; |
|---|
| 24 | default: // Approximate conditional density |
|---|
| 25 | break; |
|---|
| 26 | } |
|---|
| 27 | return vec ( 0 ); |
|---|
| 28 | } |
|---|
| 29 | |
|---|
| 30 | void merger::merge ( const epdf* g0 ) { |
|---|
| 31 | // it_file dbg ( "merger_debug.it" ); |
|---|
| 32 | |
|---|
| 33 | it_assert_debug ( rv.equal ( g0->_rv() ),"Incompatible g0" ); |
|---|
| 34 | //Empirical density - samples |
|---|
| 35 | eEmp eSmp ( rv,Ns ); |
|---|
| 36 | eSmp.set_parameters ( ones ( Ns ), g0 ); |
|---|
| 37 | Array<vec> &Smp = eSmp._samples(); //aux |
|---|
| 38 | vec &w = eSmp._w(); //aux |
|---|
| 39 | |
|---|
| 40 | mat Smp_ex =ones ( rv.count() +1,Ns ); // Extended samples for the ARX model - the last row is ones |
|---|
| 41 | for ( int i=0;i<Ns;i++ ) { set_col_part ( Smp_ex,i,Smp ( i ) );} |
|---|
| 42 | |
|---|
| 43 | // dbg << Name ( "Smp_0" ) << Smp_ex; |
|---|
| 44 | |
|---|
| 45 | // Stuff for merging |
|---|
| 46 | vec lw_src ( Ns ); |
|---|
| 47 | vec lw_mix ( Ns ); |
|---|
| 48 | mat lW=zeros ( n,Ns ); |
|---|
| 49 | vec vec0 ( 0 ); |
|---|
| 50 | |
|---|
| 51 | // Initial component in the mixture model |
|---|
| 52 | mat V0=1e-8*eye ( rv.count() +1 ); |
|---|
| 53 | ARX A0 ( RV ( "{th_r }", vec_1 ( rv.count() * ( rv.count() +1 ) ) ),\ |
|---|
| 54 | V0, rv.count() *rv.count() +5.0 ); //initial guess of Mix: zero mean, large variance |
|---|
| 55 | |
|---|
| 56 | Mix.init ( &A0, Smp_ex, Nc ); |
|---|
| 57 | //Preserve initial mixture for repetitive estimation via flattening |
|---|
| 58 | MixEF Mix_init ( Mix ); |
|---|
| 59 | |
|---|
| 60 | // ============= MAIN LOOP ================== |
|---|
| 61 | bool converged=false; |
|---|
| 62 | int niter = 0; |
|---|
| 63 | char str[100]; |
|---|
| 64 | |
|---|
| 65 | epdf* Mpred; |
|---|
| 66 | vec Mix_pdf ( Ns ); |
|---|
| 67 | while ( !converged ) { |
|---|
| 68 | //Re-estimate Mix |
|---|
| 69 | //Re-Initialize Mixture model |
|---|
| 70 | Mix.flatten ( &Mix_init ); |
|---|
| 71 | Mix.bayesB ( Smp_ex, w*Ns ); |
|---|
| 72 | Mpred = Mix.predictor ( rv ); // Allocation => must be deleted at the end!! |
|---|
| 73 | |
|---|
| 74 | // sprintf ( str,"Mpred_mean%d",niter ); |
|---|
| 75 | // dbg << Name ( str ) << Mpred->mean(); |
|---|
| 76 | |
|---|
| 77 | if ( niter<10 ) { |
|---|
| 78 | // Generate new samples |
|---|
| 79 | eSmp.set_samples ( Mpred ); |
|---|
| 80 | for ( int i=0;i<Ns;i++ ) { set_col_part ( Smp_ex,i,Smp ( i ) );} |
|---|
| 81 | |
|---|
| 82 | } |
|---|
| 83 | // sprintf ( str,"Mpdf%d",niter ); |
|---|
| 84 | // for ( int i=0;i<Ns;i++ ) {Mix_pdf ( i ) = Mix.logpred ( Smp_ex.get_col ( i ) );} |
|---|
| 85 | // dbg << Name ( str ) << Mix_pdf; |
|---|
| 86 | |
|---|
| 87 | // sprintf ( str,"Smp%d",niter ); |
|---|
| 88 | // dbg << Name ( str ) << Smp_ex; |
|---|
| 89 | |
|---|
| 90 | //Importace weighting |
|---|
| 91 | for ( int i=0;i<n;i++ ) { |
|---|
| 92 | lw_src=0.0; |
|---|
| 93 | //======== Same RVs =========== |
|---|
| 94 | //Split according to dependency in rvs |
|---|
| 95 | if ( mpdfs ( i )->_rv().count() ==rv.count() ) { |
|---|
| 96 | // no need for conditioning or marginalization |
|---|
| 97 | for ( int j=0;j<Ns; j++ ) { // Smp is Array<> => for cycle |
|---|
| 98 | lw_src ( j ) =mpdfs ( i )->_epdf().evalpdflog ( Smp ( j ) ); |
|---|
| 99 | } |
|---|
| 100 | } |
|---|
| 101 | else { |
|---|
| 102 | // compute likelihood of marginal on the conditional variable |
|---|
| 103 | if ( mpdfs ( i )->_rvc().count() >0 ) { |
|---|
| 104 | // Make marginal on rvc_i |
|---|
| 105 | epdf* tmp_marg = Mpred->marginal ( mpdfs ( i )->_rvc() ); |
|---|
| 106 | //compute vector of lw_src |
|---|
| 107 | for ( int k=0;k<Ns;k++ ) { |
|---|
| 108 | // Here val of tmp_marg = cond of mpdfs(i) ==> calling dls->get_cond |
|---|
| 109 | lw_src ( k ) += tmp_marg->evalpdflog ( dls ( i )->get_cond ( Smp ( k ) ) ); |
|---|
| 110 | } |
|---|
| 111 | delete tmp_marg; |
|---|
| 112 | |
|---|
| 113 | // sprintf ( str,"marg%d",niter ); |
|---|
| 114 | // dbg << Name ( str ) << lw_src; |
|---|
| 115 | |
|---|
| 116 | } |
|---|
| 117 | // Compute likelihood of the missing variable |
|---|
| 118 | if ( rv.count() > ( mpdfs ( i )->_rv().count() + mpdfs ( i )->_rvc().count() ) ) { |
|---|
| 119 | /////////////// |
|---|
| 120 | // There are variales unknown to mpdfs(i) : rvzs |
|---|
| 121 | mpdf* tmp_cond = Mpred->condition ( rvzs ( i ) ); |
|---|
| 122 | // Compute likelihood |
|---|
| 123 | vec lw_dbg=lw_src; |
|---|
| 124 | for ( int k= 0; k<Ns; k++ ) { |
|---|
| 125 | lw_src ( k ) += log ( |
|---|
| 126 | tmp_cond->evalcond ( |
|---|
| 127 | zdls ( i )->get_val ( Smp ( k ) ), |
|---|
| 128 | zdls ( i )->get_cond ( Smp ( k ) ) ) ); |
|---|
| 129 | } |
|---|
| 130 | delete tmp_cond; |
|---|
| 131 | } |
|---|
| 132 | // Compute likelihood of the partial source |
|---|
| 133 | for ( int k= 0; k<Ns; k++ ) { |
|---|
| 134 | mpdfs ( i )->condition ( dls ( i )->get_cond ( Smp ( k ) ) ); |
|---|
| 135 | lw_src ( k ) += mpdfs ( i )->_epdf().evalpdflog ( dls ( i )->get_val ( Smp ( k ) ) ); |
|---|
| 136 | } |
|---|
| 137 | |
|---|
| 138 | } |
|---|
| 139 | // it_assert_debug(std::isfinite(sum(lw_src)),"bad"); |
|---|
| 140 | lW.set_row ( i, lw_src ); // do not divide by mix |
|---|
| 141 | } |
|---|
| 142 | //Importance of the mixture |
|---|
| 143 | for ( int j=0;j<Ns;j++ ) { |
|---|
| 144 | lw_mix ( j ) =Mix.logpred ( Smp_ex.get_col ( j ) ); |
|---|
| 145 | } |
|---|
| 146 | // sprintf ( str,"lW%d",niter ); |
|---|
| 147 | // dbg << Name ( str ) << lW; |
|---|
| 148 | |
|---|
| 149 | w = lognorm_merge ( lW ); //merge |
|---|
| 150 | |
|---|
| 151 | // sprintf ( str,"w%d",niter ); |
|---|
| 152 | // dbg << Name ( str ) << w; |
|---|
| 153 | // sprintf ( str,"lw_m%d",niter ); |
|---|
| 154 | // dbg << Name ( str ) << lw_mix; |
|---|
| 155 | |
|---|
| 156 | //Importance weighting |
|---|
| 157 | w /=exp ( lw_mix ); // hoping that it is not numerically sensitive... |
|---|
| 158 | //renormalize |
|---|
| 159 | w /=sum ( w ); |
|---|
| 160 | |
|---|
| 161 | // sprintf ( str,"w_is_%d",niter ); |
|---|
| 162 | // dbg << Name ( str ) << w; |
|---|
| 163 | |
|---|
| 164 | // eSmp.resample(); // So that it can be used in bayes |
|---|
| 165 | // for ( int i=0;i<Ns;i++ ) { set_col_part ( Smp_ex,i,Smp ( i ) );} |
|---|
| 166 | |
|---|
| 167 | // sprintf ( str,"Smp_res%d",niter ); |
|---|
| 168 | // dbg << Name ( str ) << Smp; |
|---|
| 169 | |
|---|
| 170 | // ==== stopping rule === |
|---|
| 171 | niter++; |
|---|
| 172 | converged = ( niter>5 ); |
|---|
| 173 | } |
|---|
| 174 | |
|---|
| 175 | } |
|---|
