1 | |
---|
2 | #include "merger.h" |
---|
3 | #include "../estim/arx.h" |
---|
4 | |
---|
5 | namespace bdm |
---|
6 | { |
---|
7 | |
---|
8 | merger_base::merger_base(const Array<mpdf*> &S, bool own) { |
---|
9 | DBG = false; |
---|
10 | dbg_file = NULL; |
---|
11 | set_sources(S, own); |
---|
12 | } |
---|
13 | |
---|
14 | vec merger_base::merge_points ( mat &lW ) { |
---|
15 | int nu=lW.rows(); |
---|
16 | vec result; |
---|
17 | ivec indW; |
---|
18 | bool infexist; |
---|
19 | switch ( METHOD ) { |
---|
20 | case ARITHMETIC: |
---|
21 | result= log ( sum ( exp ( lW ) ) ); //ugly! |
---|
22 | break; |
---|
23 | case GEOMETRIC: |
---|
24 | result= sum ( lW ) /nu; |
---|
25 | break; |
---|
26 | case LOGNORMAL: |
---|
27 | vec sumlW=sum ( lW ) ; |
---|
28 | indW=find((sumlW<inf) & (sumlW>-inf)); |
---|
29 | infexist=(indW.size()<lW.cols()); |
---|
30 | vec mu; |
---|
31 | vec lam; |
---|
32 | if (infexist){ |
---|
33 | mu = sumlW(indW) /nu; //mean of logs |
---|
34 | // |
---|
35 | mat validlW=lW.get_cols(indW); |
---|
36 | lam = sum ( pow ( validlW-outer_product ( ones ( validlW.rows() ),mu ),2 ) ); |
---|
37 | } |
---|
38 | else { |
---|
39 | mu = sum ( lW ) /nu; //mean of logs |
---|
40 | lam = sum ( pow ( lW-outer_product ( ones ( lW.rows() ),mu ),2 ) ); |
---|
41 | } |
---|
42 | // |
---|
43 | double coef=0.0; |
---|
44 | vec sq2bl=sqrt ( 2*beta*lam ); //this term is everywhere |
---|
45 | switch ( nu ) { |
---|
46 | case 2: |
---|
47 | coef= ( 1-0.5*sqrt ( ( 4.0*beta-3.0 ) /beta ) ); |
---|
48 | result =coef*sq2bl + mu ; |
---|
49 | break; |
---|
50 | // case 4: == can be done similar to case 2 - is it worth it??? |
---|
51 | default: // see accompanying document merge_lognorm_derivation.lyx |
---|
52 | coef = sqrt(1-(nu+1)/(2*beta*nu)); |
---|
53 | result =log(besselk((nu-3)/2, sq2bl*coef))-log(besselk((nu-3)/2, sq2bl)) + mu; |
---|
54 | break; |
---|
55 | } |
---|
56 | break; |
---|
57 | } |
---|
58 | if (infexist){ |
---|
59 | vec tmp =-inf*ones(lW.cols()); |
---|
60 | set_subvector(tmp, indW, result); |
---|
61 | return tmp; |
---|
62 | } |
---|
63 | else {return result;} |
---|
64 | } |
---|
65 | |
---|
66 | void merger_mix::merge ( ) |
---|
67 | { |
---|
68 | Array<vec> &Smp = eSmp._samples(); //aux |
---|
69 | vec &w = eSmp._w(); //aux |
---|
70 | |
---|
71 | mat Smp_ex =ones ( dim +1,Npoints ); // Extended samples for the ARX model - the last row is ones |
---|
72 | for ( int i=0;i<Npoints;i++ ) { set_col_part ( Smp_ex,i,Smp ( i ) );} |
---|
73 | |
---|
74 | if ( DBG ) *dbg_file << Name ( "Smp_0" ) << Smp_ex; |
---|
75 | |
---|
76 | // Stuff for merging |
---|
77 | vec lw_src ( Npoints ); // weights of the ith source |
---|
78 | vec lw_mix ( Npoints ); // weights of the approximating mixture |
---|
79 | vec lw ( Npoints ); // tmp |
---|
80 | mat lW=zeros ( Nsources,Npoints ); // array of weights of all sources |
---|
81 | vec vec0 ( 0 ); |
---|
82 | |
---|
83 | //initialize importance weights |
---|
84 | lw_mix = 1.0; // assuming uniform grid density -- otherwise |
---|
85 | |
---|
86 | // Initial component in the mixture model |
---|
87 | mat V0=1e-8*eye ( dim +1 ); |
---|
88 | ARX A0; A0.set_statistics ( dim, V0 ); //initial guess of Mix: |
---|
89 | |
---|
90 | Mix.init ( &A0, Smp_ex, Ncoms ); |
---|
91 | //Preserve initial mixture for repetitive estimation via flattening |
---|
92 | MixEF Mix_init ( Mix ); |
---|
93 | |
---|
94 | // ============= MAIN LOOP ================== |
---|
95 | bool converged=false; |
---|
96 | int niter = 0; |
---|
97 | char dbg_str[100]; |
---|
98 | |
---|
99 | emix* Mpred=Mix.epredictor ( ); |
---|
100 | vec Mix_pdf ( Npoints ); |
---|
101 | while ( !converged ) |
---|
102 | { |
---|
103 | //Re-estimate Mix |
---|
104 | //Re-Initialize Mixture model |
---|
105 | Mix.flatten ( &Mix_init ); |
---|
106 | Mix.bayesB ( Smp_ex, w*Npoints ); |
---|
107 | delete Mpred; |
---|
108 | Mpred = Mix.epredictor ( ); // Allocation => must be deleted at the end!! |
---|
109 | Mpred->set_rv ( rv ); //the predictor predicts rv of this merger |
---|
110 | |
---|
111 | // This will be active only later in iterations!!! |
---|
112 | if ( 1./sum_sqr ( w ) <effss_coef*Npoints ) |
---|
113 | { |
---|
114 | // Generate new samples |
---|
115 | eSmp.set_samples ( Mpred ); |
---|
116 | for ( int i=0;i<Npoints;i++ ) |
---|
117 | { |
---|
118 | //////////// !!!!!!!!!!!!! |
---|
119 | //if ( Smp ( i ) ( 2 ) <0 ) {Smp ( i ) ( 2 ) = 0.01; } |
---|
120 | set_col_part ( Smp_ex,i,Smp ( i ) ); |
---|
121 | //Importance of the mixture |
---|
122 | //lw_mix ( i ) =Mix.logpred (Smp_ex.get_col(i) ); |
---|
123 | lw_mix ( i ) = Mpred->evallog ( Smp ( i ) ); |
---|
124 | } |
---|
125 | if ( DBG ) |
---|
126 | { |
---|
127 | cout<<"Resampling =" << 1./sum_sqr ( w ) << endl; |
---|
128 | cout << Mix._e()->mean() <<endl; |
---|
129 | cout << sum ( Smp_ex,2 ) /Npoints <<endl; |
---|
130 | cout << Smp_ex*Smp_ex.T() /Npoints << endl; |
---|
131 | } |
---|
132 | } |
---|
133 | if ( DBG ) |
---|
134 | { |
---|
135 | sprintf ( dbg_str,"Mpred_mean%d",niter ); |
---|
136 | *dbg_file << Name ( dbg_str ) << Mpred->mean(); |
---|
137 | sprintf ( dbg_str,"Mpred_var%d",niter ); |
---|
138 | *dbg_file << Name ( dbg_str ) << Mpred->variance(); |
---|
139 | |
---|
140 | |
---|
141 | sprintf ( dbg_str,"Mpdf%d",niter ); |
---|
142 | for ( int i=0;i<Npoints;i++ ) {Mix_pdf ( i ) = Mix.logpred ( Smp_ex.get_col ( i ) );} |
---|
143 | *dbg_file << Name ( dbg_str ) << Mix_pdf; |
---|
144 | |
---|
145 | sprintf ( dbg_str,"Smp%d",niter ); |
---|
146 | *dbg_file << Name ( dbg_str ) << Smp_ex; |
---|
147 | |
---|
148 | } |
---|
149 | //Importace weighting |
---|
150 | for ( int i=0;i<mpdfs.length();i++ ) |
---|
151 | { |
---|
152 | lw_src=0.0; |
---|
153 | //======== Same RVs =========== |
---|
154 | //Split according to dependency in rvs |
---|
155 | if ( mpdfs ( i )->dimension() ==dim ) |
---|
156 | { |
---|
157 | // no need for conditioning or marginalization |
---|
158 | lw_src = mpdfs ( i )->_epdf().evallog_m ( Smp ); |
---|
159 | } |
---|
160 | else |
---|
161 | { |
---|
162 | // compute likelihood of marginal on the conditional variable |
---|
163 | if ( mpdfs ( i )->dimensionc() >0 ) |
---|
164 | { |
---|
165 | // Make marginal on rvc_i |
---|
166 | epdf* tmp_marg = Mpred->marginal ( mpdfs ( i )->_rvc() ); |
---|
167 | //compute vector of lw_src |
---|
168 | for ( int k=0;k<Npoints;k++ ) |
---|
169 | { |
---|
170 | // Here val of tmp_marg = cond of mpdfs(i) ==> calling dls->get_cond |
---|
171 | lw_src ( k ) += tmp_marg->evallog ( dls ( i )->get_cond ( Smp ( k ) ) ); |
---|
172 | } |
---|
173 | delete tmp_marg; |
---|
174 | |
---|
175 | // sprintf ( str,"marg%d",niter ); |
---|
176 | // *dbg << Name ( str ) << lw_src; |
---|
177 | |
---|
178 | } |
---|
179 | // Compute likelihood of the missing variable |
---|
180 | if ( dim > ( mpdfs ( i )->dimension() + mpdfs ( i )->dimensionc() ) ) |
---|
181 | { |
---|
182 | /////////////// |
---|
183 | // There are variales unknown to mpdfs(i) : rvzs |
---|
184 | mpdf* tmp_cond = Mpred->condition ( rvzs ( i ) ); |
---|
185 | // Compute likelihood |
---|
186 | vec lw_dbg=lw_src; |
---|
187 | for ( int k= 0; k<Npoints; k++ ) |
---|
188 | { |
---|
189 | lw_src ( k ) += log ( |
---|
190 | tmp_cond->evallogcond ( |
---|
191 | zdls ( i )->pushdown ( Smp ( k ) ), |
---|
192 | zdls ( i )->get_cond ( Smp ( k ) ) ) ); |
---|
193 | if ( !std::isfinite ( lw_src ( k ) ) ) |
---|
194 | { |
---|
195 | lw_src ( k ) = -1e16; cout << "!"; |
---|
196 | } |
---|
197 | } |
---|
198 | delete tmp_cond; |
---|
199 | } |
---|
200 | // Compute likelihood of the partial source |
---|
201 | for ( int k= 0; k<Npoints; k++ ) |
---|
202 | { |
---|
203 | mpdfs ( i )->condition ( dls ( i )->get_cond ( Smp ( k ) ) ); |
---|
204 | lw_src ( k ) += mpdfs ( i )->_epdf().evallog ( dls ( i )->pushdown ( Smp ( k ) ) ); |
---|
205 | } |
---|
206 | |
---|
207 | } |
---|
208 | // it_assert_debug(std::isfinite(sum(lw_src)),"bad"); |
---|
209 | lW.set_row ( i, lw_src ); // do not divide by mix |
---|
210 | } |
---|
211 | lw = merger_base::merge_points ( lW ); //merge |
---|
212 | |
---|
213 | //Importance weighting |
---|
214 | lw -= lw_mix; // hoping that it is not numerically sensitive... |
---|
215 | w = exp ( lw-max ( lw ) ); |
---|
216 | |
---|
217 | //renormalize |
---|
218 | double sumw =sum ( w ); |
---|
219 | if ( std::isfinite ( sumw ) ) |
---|
220 | { |
---|
221 | w = w/sumw; |
---|
222 | } |
---|
223 | else |
---|
224 | { |
---|
225 | it_file itf ( "merg_err.it" ); |
---|
226 | itf << Name ( "w" ) << w; |
---|
227 | } |
---|
228 | |
---|
229 | if ( DBG ) |
---|
230 | { |
---|
231 | sprintf ( dbg_str,"lW%d",niter ); |
---|
232 | *dbg_file << Name ( dbg_str ) << lW; |
---|
233 | sprintf ( dbg_str,"w%d",niter ); |
---|
234 | *dbg_file << Name ( dbg_str ) << w; |
---|
235 | sprintf ( dbg_str,"lw_m%d",niter ); |
---|
236 | *dbg_file << Name ( dbg_str ) << lw_mix; |
---|
237 | } |
---|
238 | // ==== stopping rule === |
---|
239 | niter++; |
---|
240 | converged = ( niter>stop_niter ); |
---|
241 | } |
---|
242 | delete Mpred; |
---|
243 | // cout << endl; |
---|
244 | |
---|
245 | } |
---|
246 | |
---|
247 | // DEFAULTS FOR MERGER_BASE |
---|
248 | const MERGER_METHOD merger_base::DFLT_METHOD=LOGNORMAL; |
---|
249 | const double merger_base::DFLT_beta=1.2; |
---|
250 | // DEFAULTS FOR MERGER_MIX |
---|
251 | const int merger_mix::DFLT_Ncoms=10; |
---|
252 | const double merger_mix::DFLT_effss_coef=0.5; |
---|
253 | |
---|
254 | } |
---|