1 | /*! |
---|
2 | \file |
---|
3 | \brief Probability distributions for Mixtures of pdfs |
---|
4 | \author Vaclav Smidl. |
---|
5 | |
---|
6 | ----------------------------------- |
---|
7 | BDM++ - C++ library for Bayesian Decision Making under Uncertainty |
---|
8 | |
---|
9 | Using IT++ for numerical operations |
---|
10 | ----------------------------------- |
---|
11 | */ |
---|
12 | |
---|
13 | #ifndef MX_H |
---|
14 | #define MX_H |
---|
15 | |
---|
16 | #define LOG2 0.69314718055995 |
---|
17 | |
---|
18 | #include "libBM.h" |
---|
19 | #include "libEF.h" |
---|
20 | //#include <std> |
---|
21 | |
---|
22 | namespace bdm { |
---|
23 | |
---|
24 | //this comes first because it is used inside emix! |
---|
25 | |
---|
26 | /*! \brief Class representing ratio of two densities |
---|
27 | which arise e.g. by applying the Bayes rule. |
---|
28 | It represents density in the form: |
---|
29 | \f[ |
---|
30 | f(rv|rvc) = \frac{f(rv,rvc)}{f(rvc)} |
---|
31 | \f] |
---|
32 | where \f$ f(rvc) = \int f(rv,rvc) d\ rv \f$. |
---|
33 | |
---|
34 | In particular this type of arise by conditioning of a mixture model. |
---|
35 | |
---|
36 | At present the only supported operation is evallogcond(). |
---|
37 | */ |
---|
38 | class mratio: public mpdf { |
---|
39 | protected: |
---|
40 | //! Nominator in the form of mpdf |
---|
41 | const epdf* nom; |
---|
42 | //!Denominator in the form of epdf |
---|
43 | epdf* den; |
---|
44 | //!flag for destructor |
---|
45 | bool destroynom; |
---|
46 | //!datalink between conditional and nom |
---|
47 | datalink_m2e dl; |
---|
48 | public: |
---|
49 | //!Default constructor. By default, the given epdf is not copied! |
---|
50 | //! It is assumed that this function will be used only temporarily. |
---|
51 | mratio ( const epdf* nom0, const RV &rv, bool copy=false ) :mpdf ( ), dl ( ) { |
---|
52 | // adjust rv and rvc |
---|
53 | rvc = nom0->_rv().subt ( rv ); |
---|
54 | dimc = rvc._dsize(); |
---|
55 | ep = new epdf; |
---|
56 | ep->set_parameters(rv._dsize()); |
---|
57 | ep->set_rv(rv); |
---|
58 | |
---|
59 | //prepare data structures |
---|
60 | if ( copy ) {it_error ( "todo" ); destroynom=true; } |
---|
61 | else { nom = nom0; destroynom = false; } |
---|
62 | it_assert_debug ( rvc.length() >0,"Makes no sense to use this object!" ); |
---|
63 | |
---|
64 | // build denominator |
---|
65 | den = nom->marginal ( rvc ); |
---|
66 | dl.set_connection(rv,rvc,nom0->_rv()); |
---|
67 | }; |
---|
68 | double evallogcond ( const vec &val, const vec &cond ) { |
---|
69 | double tmp; |
---|
70 | vec nom_val ( ep->dimension() + dimc ); |
---|
71 | dl.pushup_cond ( nom_val,val,cond ); |
---|
72 | tmp = exp ( nom->evallog ( nom_val ) - den->evallog ( cond ) ); |
---|
73 | it_assert_debug ( std::isfinite ( tmp ),"Infinite value" ); |
---|
74 | return tmp; |
---|
75 | } |
---|
76 | //! Object takes ownership of nom and will destroy it |
---|
77 | void ownnom() {destroynom=true;} |
---|
78 | //! Default destructor |
---|
79 | ~mratio() {delete den; if ( destroynom ) {delete nom;}} |
---|
80 | }; |
---|
81 | |
---|
82 | /*! |
---|
83 | * \brief Mixture of epdfs |
---|
84 | |
---|
85 | Density function: |
---|
86 | \f[ |
---|
87 | f(x) = \sum_{i=1}^{n} w_{i} f_i(x), \quad \sum_{i=1}^n w_i = 1. |
---|
88 | \f] |
---|
89 | where \f$f_i(x)\f$ is any density on random variable \f$x\f$, called \a component, |
---|
90 | |
---|
91 | */ |
---|
92 | class emix : public epdf { |
---|
93 | protected: |
---|
94 | //! weights of the components |
---|
95 | vec w; |
---|
96 | //! Component (epdfs) |
---|
97 | Array<epdf*> Coms; |
---|
98 | //!Flag if owning Coms |
---|
99 | bool destroyComs; |
---|
100 | public: |
---|
101 | //!Default constructor |
---|
102 | emix ( ) : epdf ( ) {}; |
---|
103 | //! Set weights \c w and components \c Coms |
---|
104 | //!By default Coms are copied inside. Parameter \c copy can be set to false if Coms live externally. Use method ownComs() if Coms should be destroyed by the destructor. |
---|
105 | void set_parameters ( const vec &w, const Array<epdf*> &Coms, bool copy=false ); |
---|
106 | |
---|
107 | vec sample() const; |
---|
108 | vec mean() const { |
---|
109 | int i; vec mu = zeros ( dim ); |
---|
110 | for ( i = 0;i < w.length();i++ ) {mu += w ( i ) * Coms ( i )->mean(); } |
---|
111 | return mu; |
---|
112 | } |
---|
113 | vec variance() const { |
---|
114 | //non-central moment |
---|
115 | vec mom2 = zeros ( dim ); |
---|
116 | for ( int i = 0;i < w.length();i++ ) {mom2 += w ( i ) * (Coms(i)->variance() + pow ( Coms ( i )->mean(),2 )); } |
---|
117 | //central moment |
---|
118 | return mom2-pow ( mean(),2 ); |
---|
119 | } |
---|
120 | double evallog ( const vec &val ) const { |
---|
121 | int i; |
---|
122 | double sum = 0.0; |
---|
123 | for ( i = 0;i < w.length();i++ ) {sum += w ( i ) * exp ( Coms ( i )->evallog ( val ) );} |
---|
124 | if ( sum==0.0 ) {sum=std::numeric_limits<double>::epsilon();} |
---|
125 | double tmp=log ( sum ); |
---|
126 | it_assert_debug ( std::isfinite ( tmp ),"Infinite" ); |
---|
127 | return tmp; |
---|
128 | }; |
---|
129 | vec evallog_m ( const mat &Val ) const { |
---|
130 | vec x=zeros ( Val.cols() ); |
---|
131 | for ( int i = 0; i < w.length(); i++ ) { |
---|
132 | x+= w ( i ) *exp ( Coms ( i )->evallog_m ( Val ) ); |
---|
133 | } |
---|
134 | return log ( x ); |
---|
135 | }; |
---|
136 | //! Auxiliary function that returns pdflog for each component |
---|
137 | mat evallog_M ( const mat &Val ) const { |
---|
138 | mat X ( w.length(), Val.cols() ); |
---|
139 | for ( int i = 0; i < w.length(); i++ ) { |
---|
140 | X.set_row ( i, w ( i ) *exp ( Coms ( i )->evallog_m ( Val ) ) ); |
---|
141 | } |
---|
142 | return X; |
---|
143 | }; |
---|
144 | |
---|
145 | emix* marginal ( const RV &rv ) const; |
---|
146 | mratio* condition ( const RV &rv ) const; //why not mratio!! |
---|
147 | |
---|
148 | //Access methods |
---|
149 | //! returns a pointer to the internal mean value. Use with Care! |
---|
150 | vec& _w() {return w;} |
---|
151 | virtual ~emix() {if ( destroyComs ) {for ( int i=0;i<Coms.length();i++ ) {delete Coms ( i );}}} |
---|
152 | //! Auxiliary function for taking ownership of the Coms() |
---|
153 | void ownComs() {destroyComs=true;} |
---|
154 | |
---|
155 | //!access function |
---|
156 | epdf* _Coms ( int i ) {return Coms ( i );} |
---|
157 | void set_rv(const RV &rv){ |
---|
158 | epdf::set_rv(rv); |
---|
159 | for(int i=0;i<Coms.length();i++){Coms(i)->set_rv(rv);} |
---|
160 | } |
---|
161 | }; |
---|
162 | |
---|
163 | |
---|
164 | /*! |
---|
165 | * \brief Mixture of egiws |
---|
166 | |
---|
167 | */ |
---|
168 | class egiwmix : public egiw { |
---|
169 | protected: |
---|
170 | //! weights of the components |
---|
171 | vec w; |
---|
172 | //! Component (epdfs) |
---|
173 | Array<egiw*> Coms; |
---|
174 | //!Flag if owning Coms |
---|
175 | bool destroyComs; |
---|
176 | public: |
---|
177 | //!Default constructor |
---|
178 | egiwmix ( ) : egiw ( ) {}; |
---|
179 | |
---|
180 | //! Set weights \c w and components \c Coms |
---|
181 | //!By default Coms are copied inside. Parameter \c copy can be set to false if Coms live externally. Use method ownComs() if Coms should be destroyed by the destructor. |
---|
182 | void set_parameters ( const vec &w, const Array<egiw*> &Coms, bool copy=false ); |
---|
183 | |
---|
184 | //!return expected value |
---|
185 | vec mean() const; |
---|
186 | |
---|
187 | //!return a sample from the density |
---|
188 | vec sample() const; |
---|
189 | |
---|
190 | //!return the expected variance |
---|
191 | vec variance() const; |
---|
192 | |
---|
193 | // TODO!!! Defined to follow ANSI and/or for future development |
---|
194 | void mean_mat ( mat &M, mat&R ) const {}; |
---|
195 | double evallog_nn ( const vec &val ) const {return 0;}; |
---|
196 | double lognc () const {return 0;}; |
---|
197 | emix* marginal ( const RV &rv ) const; |
---|
198 | |
---|
199 | //Access methods |
---|
200 | //! returns a pointer to the internal mean value. Use with Care! |
---|
201 | vec& _w() {return w;} |
---|
202 | virtual ~egiwmix() {if ( destroyComs ) {for ( int i=0;i<Coms.length();i++ ) {delete Coms ( i );}}} |
---|
203 | //! Auxiliary function for taking ownership of the Coms() |
---|
204 | void ownComs() {destroyComs=true;} |
---|
205 | |
---|
206 | //!access function |
---|
207 | egiw* _Coms ( int i ) {return Coms ( i );} |
---|
208 | |
---|
209 | void set_rv(const RV &rv){ |
---|
210 | egiw::set_rv(rv); |
---|
211 | for(int i=0;i<Coms.length();i++){Coms(i)->set_rv(rv);} |
---|
212 | } |
---|
213 | |
---|
214 | //! Approximation of a GiW mix by a single GiW pdf |
---|
215 | egiw* approx(); |
---|
216 | }; |
---|
217 | |
---|
218 | /*! \brief Chain rule decomposition of epdf |
---|
219 | |
---|
220 | Probability density in the form of Chain-rule decomposition: |
---|
221 | \[ |
---|
222 | f(x_1,x_2,x_3) = f(x_1|x_2,x_3)f(x_2,x_3)f(x_3) |
---|
223 | \] |
---|
224 | Note that |
---|
225 | */ |
---|
226 | class mprod: public compositepdf, public mpdf { |
---|
227 | protected: |
---|
228 | //! pointers to epdfs - shortcut to mpdfs().posterior() |
---|
229 | Array<epdf*> epdfs; |
---|
230 | //! Data link for each mpdfs |
---|
231 | Array<datalink_m2m*> dls; |
---|
232 | //! dummy ep |
---|
233 | epdf dummy; |
---|
234 | public: |
---|
235 | /*!\brief Constructor from list of mFacs, |
---|
236 | */ |
---|
237 | mprod ( Array<mpdf*> mFacs ) : compositepdf ( mFacs ), mpdf (), epdfs ( n ), dls ( n ) { |
---|
238 | ep=&dummy; |
---|
239 | RV rv=getrv ( true ); |
---|
240 | set_rv ( rv );dummy.set_parameters ( rv._dsize() ); |
---|
241 | setrvc ( ep->_rv(),rvc ); |
---|
242 | // rv and rvc established = > we can link them with mpdfs |
---|
243 | for ( int i = 0;i < n;i++ ) { |
---|
244 | dls ( i ) = new datalink_m2m; |
---|
245 | dls(i)->set_connection( mpdfs ( i )->_rv(), mpdfs ( i )->_rvc(), _rv(), _rvc() ); |
---|
246 | } |
---|
247 | |
---|
248 | for ( int i=0;i<n;i++ ) { |
---|
249 | epdfs ( i ) =& ( mpdfs ( i )->_epdf() ); |
---|
250 | } |
---|
251 | }; |
---|
252 | |
---|
253 | double evallogcond ( const vec &val, const vec &cond ) { |
---|
254 | int i; |
---|
255 | double res = 0.0; |
---|
256 | for ( i = n - 1;i >= 0;i-- ) { |
---|
257 | /* if ( mpdfs(i)->_rvc().count() >0) { |
---|
258 | mpdfs ( i )->condition ( dls ( i )->get_cond ( val,cond ) ); |
---|
259 | } |
---|
260 | // add logarithms |
---|
261 | res += epdfs ( i )->evallog ( dls ( i )->pushdown ( val ) );*/ |
---|
262 | res += mpdfs ( i )->evallogcond ( |
---|
263 | dls ( i )->pushdown ( val ), |
---|
264 | dls ( i )->get_cond ( val, cond ) |
---|
265 | ); |
---|
266 | } |
---|
267 | return res; |
---|
268 | } |
---|
269 | //TODO smarter... |
---|
270 | vec samplecond ( const vec &cond ) { |
---|
271 | //! Ugly hack to help to discover if mpfs are not in proper order. Correct solution = check that explicitely. |
---|
272 | vec smp= std::numeric_limits<double>::infinity() * ones ( ep->dimension() ); |
---|
273 | vec smpi; |
---|
274 | // Hard assumption here!!! We are going backwards, to assure that samples that are needed from smp are already generated! |
---|
275 | for ( int i = ( n - 1 );i >= 0;i-- ) { |
---|
276 | if ( mpdfs ( i )->dimensionc() ) { |
---|
277 | mpdfs ( i )->condition ( dls ( i )->get_cond ( smp ,cond ) ); // smp is val here!! |
---|
278 | } |
---|
279 | smpi = epdfs ( i )->sample(); |
---|
280 | // copy contribution of this pdf into smp |
---|
281 | dls ( i )->pushup ( smp, smpi ); |
---|
282 | // add ith likelihood contribution |
---|
283 | } |
---|
284 | return smp; |
---|
285 | } |
---|
286 | mat samplecond ( const vec &cond, int N ) { |
---|
287 | mat Smp ( dimension(),N ); |
---|
288 | for ( int i=0;i<N;i++ ) {Smp.set_col ( i,samplecond ( cond ) );} |
---|
289 | return Smp; |
---|
290 | } |
---|
291 | |
---|
292 | ~mprod() {}; |
---|
293 | }; |
---|
294 | |
---|
295 | //! Product of independent epdfs. For dependent pdfs, use mprod. |
---|
296 | class eprod: public epdf { |
---|
297 | protected: |
---|
298 | //! Components (epdfs) |
---|
299 | Array<const epdf*> epdfs; |
---|
300 | //! Array of indeces |
---|
301 | Array<datalink*> dls; |
---|
302 | public: |
---|
303 | eprod () : epdfs ( 0 ),dls ( 0 ) {}; |
---|
304 | void set_parameters ( const Array<const epdf*> &epdfs0, bool named=true ) { |
---|
305 | epdfs=epdfs0;//.set_length ( epdfs0.length() ); |
---|
306 | dls.set_length ( epdfs.length() ); |
---|
307 | |
---|
308 | bool independent=true; |
---|
309 | if ( named ) { |
---|
310 | for ( int i=0;i<epdfs.length();i++ ) { |
---|
311 | independent=rv.add ( epdfs ( i )->_rv() ); |
---|
312 | it_assert_debug ( independent==true, "eprod:: given components are not independent." ); |
---|
313 | } |
---|
314 | dim=rv._dsize(); |
---|
315 | } |
---|
316 | else { |
---|
317 | dim =0; for ( int i=0;i<epdfs.length();i++ ) { |
---|
318 | dim+=epdfs ( i )->dimension(); |
---|
319 | } |
---|
320 | } |
---|
321 | // |
---|
322 | int cumdim=0; |
---|
323 | int dimi=0; |
---|
324 | int i; |
---|
325 | for ( i=0;i<epdfs.length();i++ ) { |
---|
326 | dls ( i ) = new datalink; |
---|
327 | if ( named ) {dls ( i )->set_connection ( epdfs ( i )->_rv() , rv );} |
---|
328 | else { |
---|
329 | dimi = epdfs ( i )->dimension(); |
---|
330 | dls ( i )->set_connection ( dimi, dim, linspace ( cumdim,cumdim+dimi-1 ) ); |
---|
331 | cumdim+=dimi; |
---|
332 | } |
---|
333 | } |
---|
334 | } |
---|
335 | |
---|
336 | vec mean() const { |
---|
337 | vec tmp ( dim ); |
---|
338 | for ( int i=0;i<epdfs.length();i++ ) { |
---|
339 | vec pom = epdfs ( i )->mean(); |
---|
340 | dls ( i )->pushup ( tmp, pom ); |
---|
341 | } |
---|
342 | return tmp; |
---|
343 | } |
---|
344 | vec variance() const { |
---|
345 | vec tmp ( dim ); //second moment |
---|
346 | for ( int i=0;i<epdfs.length();i++ ) { |
---|
347 | vec pom = epdfs ( i )->mean(); |
---|
348 | dls ( i )->pushup ( tmp, pow ( pom,2 ) ); |
---|
349 | } |
---|
350 | return tmp-pow ( mean(),2 ); |
---|
351 | } |
---|
352 | vec sample() const { |
---|
353 | vec tmp ( dim ); |
---|
354 | for ( int i=0;i<epdfs.length();i++ ) { |
---|
355 | vec pom = epdfs ( i )->sample(); |
---|
356 | dls ( i )->pushup ( tmp, pom ); |
---|
357 | } |
---|
358 | return tmp; |
---|
359 | } |
---|
360 | double evallog ( const vec &val ) const { |
---|
361 | double tmp=0; |
---|
362 | for ( int i=0;i<epdfs.length();i++ ) { |
---|
363 | tmp+=epdfs ( i )->evallog ( dls ( i )->pushdown ( val ) ); |
---|
364 | } |
---|
365 | it_assert_debug ( std::isfinite ( tmp ),"Infinite" ); |
---|
366 | return tmp; |
---|
367 | } |
---|
368 | //!access function |
---|
369 | const epdf* operator () ( int i ) const {it_assert_debug ( i<epdfs.length(),"wrong index" );return epdfs ( i );} |
---|
370 | |
---|
371 | //!Destructor |
---|
372 | ~eprod() {for ( int i=0;i<epdfs.length();i++ ) {delete dls ( i );}} |
---|
373 | }; |
---|
374 | |
---|
375 | |
---|
376 | /*! \brief Mixture of mpdfs with constant weights, all mpdfs are of equal type |
---|
377 | |
---|
378 | */ |
---|
379 | class mmix : public mpdf { |
---|
380 | protected: |
---|
381 | //! Component (epdfs) |
---|
382 | Array<mpdf*> Coms; |
---|
383 | //!Internal epdf |
---|
384 | emix Epdf; |
---|
385 | public: |
---|
386 | //!Default constructor |
---|
387 | mmix ( ) : mpdf ( ), Epdf () {ep = &Epdf;}; |
---|
388 | //! Set weights \c w and components \c R |
---|
389 | void set_parameters ( const vec &w, const Array<mpdf*> &Coms ) { |
---|
390 | Array<epdf*> Eps ( Coms.length() ); |
---|
391 | |
---|
392 | for ( int i = 0;i < Coms.length();i++ ) { |
---|
393 | Eps ( i ) = & ( Coms ( i )->_epdf() ); |
---|
394 | } |
---|
395 | Epdf.set_parameters ( w, Eps ); |
---|
396 | }; |
---|
397 | |
---|
398 | void condition ( const vec &cond ) { |
---|
399 | for ( int i = 0;i < Coms.length();i++ ) {Coms ( i )->condition ( cond );} |
---|
400 | }; |
---|
401 | }; |
---|
402 | |
---|
403 | } |
---|
404 | #endif //MX_H |
---|