[646] | 1 | |
---|
| 2 | |
---|
| 3 | #include <limits> |
---|
| 4 | #include "arx.h" |
---|
| 5 | |
---|
| 6 | |
---|
| 7 | #include <fstream> |
---|
| 8 | #include<iostream> |
---|
| 9 | #include<iomanip> |
---|
| 10 | |
---|
| 11 | namespace bdm { |
---|
| 12 | |
---|
[684] | 13 | |
---|
[646] | 14 | |
---|
[684] | 15 | |
---|
| 16 | /* |
---|
[646] | 17 | struct str_aux { |
---|
| 18 | vec d0; |
---|
| 19 | double nu0; |
---|
| 20 | mat L0; |
---|
| 21 | mat L; |
---|
| 22 | vec d; |
---|
| 23 | double nu; |
---|
| 24 | ivec strL; // Current structure of L and d |
---|
| 25 | ivec strRgr; // Structure elements currently inside regressor (after regressand) |
---|
| 26 | ivec strMis; // structure elements, that are currently outside regressor (before regressand) |
---|
| 27 | int posit1; // regressand position |
---|
| 28 | int nbits; // number of bits available in double |
---|
[684] | 29 | bvec bitstr; |
---|
[646] | 30 | double loglik; // loglikelihood |
---|
| 31 | }; |
---|
[684] | 32 | |
---|
[646] | 33 | */ |
---|
| 34 | |
---|
[684] | 35 | |
---|
| 36 | /* bvec str_bitset( bvec in,ivec ns,int nbits){ |
---|
| 37 | //int index, bitindex,n; |
---|
| 38 | bvec out = in; |
---|
| 39 | int n; |
---|
| 40 | |
---|
| 41 | for (int i = 0; i < ns.length(); i++){ |
---|
| 42 | n = ns(i); |
---|
| 43 | out(n-2) = 1; |
---|
| 44 | cout << out; |
---|
| 45 | |
---|
| 46 | } |
---|
| 47 | |
---|
| 48 | return out; |
---|
| 49 | |
---|
| 50 | }*/ |
---|
| 51 | |
---|
| 52 | |
---|
| 53 | void str_bitset( bvec &out,ivec ns,int nbits){ |
---|
| 54 | for (int i = 0; i < ns.length(); i++){ |
---|
| 55 | out(ns(i)-2) = 1; |
---|
| 56 | } |
---|
[646] | 57 | } |
---|
| 58 | |
---|
| 59 | |
---|
| 60 | |
---|
| 61 | |
---|
| 62 | |
---|
| 63 | |
---|
[684] | 64 | double seloglik1(str_aux in){ |
---|
[646] | 65 | // This is the loglikelihood (non-constant part) - this should be used in |
---|
| 66 | // frequent computation |
---|
[684] | 67 | int len = length(in.d); |
---|
| 68 | int p1 = in.posit1 - 1; |
---|
| 69 | |
---|
| 70 | double i1 = -0.5*in.nu *log(in.d(p1)) -0.5*sum(log(in.d.right(len - p1 -1))); |
---|
| 71 | double i0 = -0.5*in.nu0*log(in.d0(p1)) -0.5*sum(log(in.d0.right(len - p1 -1))); |
---|
| 72 | return i1-i0; |
---|
| 73 | //DEBUGGing print: |
---|
| 74 | //fprintf('SELOGLIK1: str=%s loglik=%g\n', strPrintstr(in), l);*/ |
---|
| 75 | |
---|
| 76 | } |
---|
[646] | 77 | |
---|
| 78 | |
---|
[684] | 79 | void sedydr(mat &r,mat &f,double &Dr,double &Df,int R/*,int jl,int jh ,mat &rout, mat &fout, double &Drout, double &Dfoutint &kr*/){ |
---|
| 80 | /*SEDYDR dyadic reduction, performs transformation of sum of 2 dyads |
---|
| 81 | % |
---|
| 82 | % [rout, fout, Drout, Dfout, kr] = sedydr(r,f,Dr,Df,R,jl,jh); |
---|
| 83 | % [rout, fout, Drout, Dfout] = sedydr(r,f,Dr,Df,R); |
---|
| 84 | % |
---|
| 85 | % Description: dyadic reduction, performs transformation of sum of |
---|
| 86 | % 2 dyads r*Dr*r'+ f*Df*f' so that the element of r pointed by R is zeroed |
---|
| 87 | % |
---|
| 88 | % r : column vector of reduced dyad |
---|
| 89 | % f : column vector of reducing dyad |
---|
| 90 | % Dr : scalar with weight of reduced dyad |
---|
| 91 | % Df : scalar with weight of reducing dyad |
---|
| 92 | % R : scalar number giving 1 based index to the element of r, |
---|
| 93 | % which is to be reduced to |
---|
| 94 | % zero; the corresponding element of f is assumed to be 1. |
---|
| 95 | % jl : lower index of the range within which the dyads are |
---|
| 96 | % modified (can be omitted, then everything is updated) |
---|
| 97 | % jh : upper index of the range within which the dyads are |
---|
| 98 | % modified (can be omitted then everything is updated) |
---|
| 99 | % rout,fout,Drout,dfout : resulting two dyads |
---|
| 100 | % kr : coefficient used in the transformation of r |
---|
| 101 | % rnew = r + kr*f |
---|
| 102 | % |
---|
| 103 | % Description: dyadic reduction, performs transformation of sum of |
---|
| 104 | % 2 dyads r*Dr*r'+ f*Df*f' so that the element of r indexed by R is zeroed |
---|
| 105 | % Remark1: Constant mzero means machine zero and should be modified |
---|
| 106 | % according to the precision of particular machine |
---|
| 107 | % Remark2: jl and jh are, in fact, obsolete. It takes longer time to |
---|
| 108 | % compute them compared to plain version. The reason is that we |
---|
| 109 | % are doing vector operations in m-file. Other reason is that |
---|
| 110 | % we need to copy whole vector anyway. It can save half of time for |
---|
| 111 | % c-file, if you use it correctly. (please do tests) |
---|
| 112 | % |
---|
| 113 | % Note: naming: |
---|
| 114 | % se = structure estimation |
---|
| 115 | % dydr = dyadic reduction |
---|
| 116 | % |
---|
| 117 | % Original Fortran design: V. Peterka 17-7-89 |
---|
| 118 | % Modified for c-language: probably R. Kulhavy |
---|
| 119 | % Modified for m-language: L. Tesar 2/2003 |
---|
| 120 | % Updated: Feb 2003 |
---|
| 121 | % Project: post-ProDaCTool |
---|
| 122 | % Reference: none*/ |
---|
| 123 | |
---|
| 124 | /*if nargin<6; |
---|
| 125 | update_whole=1; |
---|
| 126 | else |
---|
| 127 | update_whole=0; |
---|
| 128 | end;*/ |
---|
| 129 | |
---|
| 130 | double mzero = 1e-32; |
---|
| 131 | |
---|
| 132 | if (Dr<mzero){ |
---|
| 133 | Dr=0; |
---|
[646] | 134 | } |
---|
[684] | 135 | |
---|
| 136 | double r0 = r(R,0); |
---|
| 137 | double kD = Df; |
---|
| 138 | double kr = r0 * Dr; |
---|
[646] | 139 | |
---|
| 140 | |
---|
[684] | 141 | Df = kD + r0 * kr; |
---|
| 142 | |
---|
| 143 | if (Df > mzero){ |
---|
| 144 | kD = kD / Df; |
---|
| 145 | kr = kr / Df; |
---|
| 146 | }else{ |
---|
| 147 | kD = 1; |
---|
| 148 | kr = 0; |
---|
| 149 | } |
---|
| 150 | |
---|
| 151 | Dr = Dr * kD; |
---|
| 152 | |
---|
[646] | 153 | // Try to uncomment marked stuff (*) if in numerical problems, but I don't |
---|
| 154 | // think it can make any difference for normal healthy floating-point unit |
---|
| 155 | //if update_whole; |
---|
[684] | 156 | r = r - r0*f; |
---|
[646] | 157 | // rout(R) = 0; // * could be needed for some nonsense cases(or numeric reasons?), normally not |
---|
[684] | 158 | f = f + kr*r; |
---|
[646] | 159 | // fout(R) = 1; // * could be needed for some nonsense cases(or numeric reasons?), normally not |
---|
[684] | 160 | /*else; |
---|
| 161 | rout = r; |
---|
| 162 | fout = f; |
---|
| 163 | rout(jl:jh) = r(jl:jh) - r0 * f(jl:jh); |
---|
| 164 | rout(R) = 0; |
---|
| 165 | fout(jl:jh) = f(jl:jh) + kr * rout(jl:jh); |
---|
| 166 | end;*/ |
---|
| 167 | } |
---|
[646] | 168 | |
---|
| 169 | |
---|
[684] | 170 | |
---|
| 171 | /*mat*/ void seswapudl(mat &L, vec &d , int i/*, vec &dout*/){ |
---|
| 172 | /*%SESWAPUDL swaps information matrix in decomposition V=L^T diag(d) L |
---|
| 173 | % |
---|
| 174 | % [Lout, dout] = seswapudl(L,d,i); |
---|
| 175 | % |
---|
| 176 | % L : lower triangular matrix with 1's on diagonal of the decomposistion |
---|
| 177 | % d : diagonal vector of diagonal matrix of the decomposition |
---|
| 178 | % i : index of line to be swapped with the next one |
---|
| 179 | % Lout : output lower triangular matrix |
---|
| 180 | % dout : output diagional vector of diagonal matrix D |
---|
| 181 | % |
---|
| 182 | % Description: |
---|
| 183 | % Lout' * diag(dout) * Lout = P(i,i+1) * L' * diag(d) * L * P(i,i+1); |
---|
| 184 | % |
---|
| 185 | % Where permutation matrix P(i,j) permutates columns if applied from the |
---|
| 186 | % right and line if applied from the left. |
---|
| 187 | % |
---|
| 188 | % Note: naming: |
---|
| 189 | % se = structure estimation |
---|
| 190 | % lite = light, simple |
---|
| 191 | % udl = U*D*L, or more precisely, L'*D*L, also called as ld |
---|
| 192 | % |
---|
| 193 | % Design : L. Tesar |
---|
| 194 | % Updated : Feb 2003 |
---|
| 195 | % Project : post-ProDaCTool |
---|
| 196 | % Reference: sedydr*/ |
---|
| 197 | |
---|
| 198 | int j = i+1; |
---|
| 199 | |
---|
| 200 | double pomd = d(i); |
---|
| 201 | d(i) = d(j); |
---|
| 202 | d(j) = pomd; |
---|
| 203 | |
---|
| 204 | /*vec pomL = L.get_row(i); |
---|
| 205 | L.set_row(i, L.get_row(j)); |
---|
| 206 | L.set_row(j,pomL);*/ |
---|
| 207 | |
---|
| 208 | L.swap_rows(i,j); |
---|
| 209 | L.swap_cols(i,j); |
---|
[646] | 210 | |
---|
[684] | 211 | /*pomL = L.get_col(i); |
---|
| 212 | L.set_col(i, L.get_col(j)); |
---|
| 213 | L.set_col(j,pomL);*/ |
---|
| 214 | |
---|
[646] | 215 | //% We must be working with LINES of matrix L ! |
---|
[684] | 216 | |
---|
[646] | 217 | |
---|
| 218 | |
---|
[684] | 219 | mat r = L.get_row(i); |
---|
| 220 | r = r.transpose(); |
---|
| 221 | r = r.transpose(); |
---|
[646] | 222 | //???????????????? |
---|
[684] | 223 | mat f = L.get_row(j); |
---|
| 224 | f = f.transpose(); |
---|
| 225 | f = f.transpose(); |
---|
[646] | 226 | |
---|
| 227 | |
---|
| 228 | |
---|
| 229 | |
---|
[684] | 230 | double Dr = d(i); |
---|
| 231 | double Df = d(j); |
---|
| 232 | |
---|
| 233 | sedydr(r, f, Dr, Df, j); |
---|
| 234 | |
---|
[646] | 235 | |
---|
| 236 | |
---|
[684] | 237 | double r0 = r(i,0); |
---|
| 238 | Dr = Dr*r0*r0; |
---|
| 239 | r = r/r0; |
---|
[646] | 240 | |
---|
| 241 | |
---|
| 242 | |
---|
[684] | 243 | mat pom_mat = r.transpose(); |
---|
| 244 | L.set_row(i, pom_mat.get_row(0)); |
---|
| 245 | pom_mat = f.transpose(); |
---|
| 246 | L.set_row(j, pom_mat.get_row(0)); |
---|
[646] | 247 | |
---|
[684] | 248 | d(i) = Dr; |
---|
| 249 | d(j) = Df; |
---|
| 250 | |
---|
| 251 | L(i,i) = 1; |
---|
| 252 | L(j,j) = 1; |
---|
| 253 | |
---|
| 254 | |
---|
| 255 | } |
---|
| 256 | |
---|
[646] | 257 | |
---|
[684] | 258 | void str_bitres(bvec &out,ivec ns,int nbits){ |
---|
| 259 | |
---|
| 260 | |
---|
| 261 | for (int i = 0; i < ns.length(); i++){ |
---|
| 262 | out(ns(i)-2) = 0; |
---|
| 263 | } |
---|
| 264 | |
---|
| 265 | |
---|
| 266 | } |
---|
| 267 | |
---|
| 268 | str_aux sestrremove(str_aux in,ivec removed_elements){ |
---|
[646] | 269 | //% Removes elements from regressor |
---|
[684] | 270 | int n_strL = length(in.strL); |
---|
| 271 | str_aux out = in; |
---|
| 272 | for (int i = 0; i < removed_elements.length();i++){ |
---|
| 273 | |
---|
| 274 | int f = removed_elements(i); |
---|
| 275 | int posit1 = (find(out.strL==1))(0); |
---|
| 276 | int positf = (find(out.strL==f))(0); |
---|
| 277 | int pom_strL; |
---|
| 278 | for (int g = positf-1; g >posit1 -1; g--) { |
---|
| 279 | //% BEGIN: We are swapping g and g+1 NOW!!!! |
---|
| 280 | seswapudl(out.L, out.d, g); |
---|
| 281 | seswapudl(out.L0, out.d0, g); |
---|
| 282 | |
---|
| 283 | pom_strL = out.strL(g); |
---|
| 284 | out.strL(g)= out.strL(g+1); |
---|
| 285 | out.strL(g+1) = pom_strL; |
---|
| 286 | |
---|
| 287 | //% END |
---|
| 288 | } |
---|
| 289 | } |
---|
| 290 | out.posit1 = (find(out.strL==1))(0)+1; |
---|
| 291 | out.strRgr = out.strL.right(n_strL - out.posit1); |
---|
| 292 | out.strMis = out.strL.left(out.posit1-1); |
---|
| 293 | str_bitres(out.bitstr,removed_elements,out.nbits); |
---|
| 294 | out.loglik = seloglik1(out); |
---|
| 295 | |
---|
| 296 | return out; |
---|
| 297 | } |
---|
| 298 | |
---|
[646] | 299 | |
---|
[684] | 300 | ivec setdiff(ivec a, ivec b){ |
---|
| 301 | ivec pos; |
---|
| 302 | |
---|
| 303 | for (int i = 0; i < b.length(); i++){ |
---|
| 304 | pos = find(a==b(i)); |
---|
| 305 | for (int j = 0; j < pos.length(); j++){ |
---|
| 306 | a.del(pos(j)-j); |
---|
[646] | 307 | } |
---|
| 308 | } |
---|
[684] | 309 | return a; |
---|
| 310 | } |
---|
[646] | 311 | |
---|
| 312 | |
---|
[684] | 313 | |
---|
[646] | 314 | /* |
---|
[684] | 315 | |
---|
[646] | 316 | Array<str_aux> add_new(Array<str_aux> global_best,str_aux newone,int nbest){ |
---|
| 317 | // Eventually add to global best, but do not go over nbest values |
---|
| 318 | // Also avoids repeating things, which makes this function awfully slow |
---|
[684] | 319 | |
---|
[646] | 320 | Array<str_aux> global_best_out; |
---|
| 321 | if (global_best.length() >= nbest){ |
---|
| 322 | //logliks = [global_best.loglik]; |
---|
[684] | 323 | |
---|
| 324 | vec logliks(1); |
---|
[646] | 325 | logliks(0) = global_best(0).loglik; |
---|
| 326 | for (int j = 1; j < global_best.length(); j++) |
---|
| 327 | logliks = concat(logliks, global_best(j).loglik); |
---|
[684] | 328 | |
---|
[646] | 329 | int i, addit; |
---|
| 330 | double loglik = min(logliks, i); |
---|
| 331 | global_best_out = global_best; |
---|
| 332 | if (loglik < newone.loglik){ |
---|
| 333 | // if ~any(logliks == new.loglik); |
---|
| 334 | addit=1; |
---|
[684] | 335 | |
---|
| 336 | |
---|
| 337 | |
---|
[646] | 338 | if (newone.bitstr.length() == 1) { |
---|
| 339 | for (int j = 0; j < global_best.length(); j++){ |
---|
| 340 | for(int i = 0; i < global_best(j).bitstr.length(); i++){ |
---|
| 341 | |
---|
| 342 | if (newone.bitstr(0) == global_best(j).bitstr(i)){ |
---|
| 343 | addit = 0; |
---|
| 344 | break; |
---|
| 345 | } |
---|
[684] | 346 | } |
---|
| 347 | } |
---|
[646] | 348 | } |
---|
| 349 | if (addit){ |
---|
| 350 | global_best_out(i) = newone; |
---|
| 351 | // DEBUGging print: |
---|
| 352 | // fprintf('ADDED structure, add_new: %s, loglik=%g\n', strPrintstr(new), new.loglik); |
---|
[684] | 353 | } |
---|
| 354 | } |
---|
[646] | 355 | } |
---|
| 356 | else |
---|
| 357 | global_best_out = concat(global_best, newone); |
---|
| 358 | |
---|
| 359 | return global_best_out; |
---|
[684] | 360 | |
---|
[646] | 361 | } |
---|
[684] | 362 | |
---|
[646] | 363 | */ |
---|
[684] | 364 | |
---|
| 365 | void add_new(Array<str_aux> &global_best,str_aux newone,int nbest){ |
---|
[646] | 366 | // Eventually add to global best, but do not go over nbest values |
---|
| 367 | // Also avoids repeating things, which makes this function awfully slow |
---|
[684] | 368 | |
---|
| 369 | int addit, i = 0; |
---|
| 370 | if (global_best.length() >= nbest){ |
---|
| 371 | //logliks = [global_best.loglik]; |
---|
| 372 | |
---|
| 373 | |
---|
| 374 | for (int j = 1; j < global_best.length(); j++){ |
---|
| 375 | if (global_best(j).loglik < global_best(i).loglik) { |
---|
| 376 | i = j; |
---|
| 377 | } |
---|
| 378 | } |
---|
| 379 | |
---|
| 380 | if (global_best(i).loglik < newone.loglik){ |
---|
| 381 | // if ~any(logliks == new.loglik); |
---|
| 382 | addit=1; |
---|
| 383 | |
---|
| 384 | |
---|
| 385 | //???????????????????????????????????????????? |
---|
| 386 | if (newone.bitstr.length() == 1) { |
---|
| 387 | for (int j = 0; j < global_best.length(); j++){ |
---|
| 388 | for(int i = 0; i < global_best(j).bitstr.length(); i++){ |
---|
[646] | 389 | |
---|
[684] | 390 | if (newone.bitstr(0) == global_best(j).bitstr(i)){ |
---|
| 391 | addit = 0; |
---|
| 392 | break; |
---|
| 393 | } |
---|
| 394 | } |
---|
| 395 | } |
---|
| 396 | } |
---|
| 397 | |
---|
| 398 | |
---|
| 399 | //????????????????????????????????????????????????? |
---|
| 400 | |
---|
| 401 | if (addit){ |
---|
| 402 | global_best(i) = newone; |
---|
| 403 | // DEBUGging print: |
---|
| 404 | // fprintf('ADDED structure, add_new: %s, loglik=%g\n', strPrintstr(new), new.loglik); |
---|
| 405 | } |
---|
| 406 | } |
---|
| 407 | } |
---|
| 408 | else |
---|
| 409 | global_best = concat(global_best, newone); |
---|
[646] | 410 | |
---|
| 411 | } |
---|
[684] | 412 | |
---|
| 413 | |
---|
| 414 | |
---|
| 415 | |
---|
| 416 | |
---|
| 417 | |
---|
| 418 | str_aux sestrinsert(str_aux in,ivec inserted_elements){ |
---|
[646] | 419 | // Moves elements into regressor |
---|
[684] | 420 | int n_strL = in.strL.length(); |
---|
| 421 | str_aux out = in; |
---|
| 422 | for (int j = 0;j < inserted_elements.length(); j++){ |
---|
| 423 | int f = inserted_elements(j); |
---|
| 424 | int posit1 = (find(out.strL==1))(0); |
---|
| 425 | int positf = (find(out.strL==f))(0); |
---|
| 426 | for (int g = positf; g <= posit1-1; g++ ){ |
---|
| 427 | |
---|
| 428 | // BEGIN: We are swapping g and g+1 NOW!!!! |
---|
| 429 | seswapudl(out.L, out.d, g); |
---|
| 430 | seswapudl(out.L0, out.d0, g); |
---|
| 431 | |
---|
| 432 | |
---|
| 433 | int pom_strL = out.strL(g); |
---|
| 434 | out.strL(g)= out.strL(g+1); |
---|
| 435 | out.strL(g+1) = pom_strL; |
---|
| 436 | |
---|
| 437 | // END |
---|
| 438 | } |
---|
| 439 | } |
---|
[646] | 440 | |
---|
[684] | 441 | out.posit1 = (find(out.strL==1))(0)+1; |
---|
| 442 | out.strRgr = out.strL.right(n_strL - out.posit1); |
---|
| 443 | out.strMis = out.strL.left(out.posit1-1); |
---|
| 444 | str_bitset(out.bitstr,inserted_elements,out.nbits); |
---|
| 445 | |
---|
| 446 | out.loglik = seloglik1(out); |
---|
| 447 | |
---|
| 448 | |
---|
| 449 | |
---|
| 450 | return out; |
---|
| 451 | |
---|
| 452 | } |
---|
| 453 | |
---|
| 454 | double seloglik2(str_aux in){ |
---|
[646] | 455 | // This is the loglikelihood (constant part) - this should be added to |
---|
| 456 | // everything at the end. It needs some computation, so it is useless to |
---|
| 457 | // make it for all the stuff |
---|
[684] | 458 | double logpi = log(pi); |
---|
| 459 | |
---|
| 460 | double i1 = lgamma(in.nu /2) - 0.5*in.nu *logpi; |
---|
| 461 | double i0 = lgamma(in.nu0/2) - 0.5*in.nu0*logpi; |
---|
| 462 | return i1-i0; |
---|
| 463 | } |
---|
[646] | 464 | |
---|
[684] | 465 | |
---|
[646] | 466 | |
---|
[684] | 467 | |
---|
| 468 | ivec straux1(ldmat Ld, double nu, ldmat Ld0, double nu0, ivec belief, int nbest, int max_nrep, double lambda, int order_k, Array<str_aux> &rgrsout/*, stat &statistics*/){ |
---|
[646] | 469 | // see utia_legacy/ticket_12/ implementation and str_test.m |
---|
| 470 | |
---|
[684] | 471 | const mat &L = Ld._L(); |
---|
| 472 | const vec &d = Ld._D(); |
---|
[646] | 473 | |
---|
[684] | 474 | const mat &L0 = Ld0._L(); |
---|
| 475 | const vec &d0 = Ld0._D(); |
---|
[646] | 476 | |
---|
[684] | 477 | int n_data = d.length(); |
---|
[646] | 478 | |
---|
[684] | 479 | ivec belief_out = find(belief==4)+2; // we are avoiding to put this into regressor |
---|
| 480 | ivec belief_in = find(belief==1)+2; // we are instantly keeping this in regressor |
---|
[646] | 481 | |
---|
| 482 | |
---|
[684] | 483 | str_aux full; |
---|
[646] | 484 | |
---|
[684] | 485 | full.d0 = d0; |
---|
| 486 | full.nu0 = nu0; |
---|
| 487 | full.L0 = L0; |
---|
| 488 | full.L = L; |
---|
| 489 | full.d = d; |
---|
[646] | 490 | |
---|
| 491 | |
---|
| 492 | |
---|
[684] | 493 | /*full.d0 = "0.012360650875200 0.975779169502626 1.209840558439000"; |
---|
| 494 | |
---|
| 495 | full.L0 = "1.0000 0 0;" |
---|
| 496 | "0.999427690134298 1.0000 0;" |
---|
| 497 | "0.546994424043659 0.534335486953833 1.0000"; |
---|
[646] | 498 | |
---|
| 499 | |
---|
[684] | 500 | full.L = "1.0000 0 0;" |
---|
| 501 | "0.364376353850780 1.0000 0;" |
---|
| 502 | "1.222141096674815 1.286534510946323 1.0000"; |
---|
[646] | 503 | |
---|
[684] | 504 | full.d = "0.001610356837691 3.497566869589465 3.236640487818002"; |
---|
| 505 | */ |
---|
[646] | 506 | |
---|
[684] | 507 | fstream F; |
---|
[646] | 508 | |
---|
| 509 | |
---|
| 510 | |
---|
| 511 | |
---|
[684] | 512 | F.open("soubor3.txt", ios::in); |
---|
| 513 | F << setiosflags(ios::scientific); |
---|
| 514 | F << setprecision(16); |
---|
| 515 | F.flags(0x1); |
---|
[646] | 516 | |
---|
| 517 | |
---|
[684] | 518 | for (int i = 0; i < n_data ; i++){ |
---|
| 519 | F >> full.d0(i); |
---|
| 520 | } |
---|
[646] | 521 | |
---|
| 522 | |
---|
| 523 | |
---|
[684] | 524 | for (int i =0; i < n_data; i++){ |
---|
| 525 | for (int j = 0 ; j < n_data ; j++){ |
---|
| 526 | F >> full.L0(j,i); |
---|
| 527 | } |
---|
| 528 | } |
---|
[646] | 529 | |
---|
[684] | 530 | for (int i = 0; i < n_data ; i++){ |
---|
| 531 | F >> full.d(i); |
---|
| 532 | } |
---|
[646] | 533 | |
---|
| 534 | |
---|
[684] | 535 | for (int i =0; i < n_data; i++){ |
---|
| 536 | for (int j = 0 ; j < n_data ; j++){ |
---|
| 537 | F >> full.L(j,i); |
---|
| 538 | } |
---|
| 539 | } |
---|
[646] | 540 | |
---|
| 541 | |
---|
[684] | 542 | full.nu0 = nu0; |
---|
| 543 | full.nu = nu; |
---|
| 544 | full.strL = linspace(1,n_data); |
---|
| 545 | full.strRgr = linspace(2,n_data); |
---|
| 546 | full.strMis = ivec(0); |
---|
| 547 | full.posit1 = 1; |
---|
| 548 | full.bitstr.set_size(n_data-1); |
---|
| 549 | full.bitstr.clear(); |
---|
| 550 | str_bitset(full.bitstr,full.strRgr,full.nbits); |
---|
[646] | 551 | //full.nbits = std::numeric_lim its<double>::digits-1; // number of bits available in double |
---|
[684] | 552 | /*bvec in(n_data-1); |
---|
| 553 | in.clear(); |
---|
| 554 | full.bitstr = str_bitset(in,full.strRgr,full.nbits);*/ |
---|
[646] | 555 | |
---|
| 556 | |
---|
| 557 | |
---|
| 558 | |
---|
[684] | 559 | full.loglik = seloglik1(full); // % loglikelihood |
---|
| 560 | |
---|
[646] | 561 | |
---|
| 562 | |
---|
| 563 | |
---|
| 564 | |
---|
| 565 | //% construct full and empty structure |
---|
[684] | 566 | full = sestrremove(full,belief_out); |
---|
| 567 | str_aux empty = sestrremove(full,setdiff(full.strRgr,belief_in)); |
---|
| 568 | |
---|
[646] | 569 | //% stopping rule calculation: |
---|
| 570 | |
---|
| 571 | |
---|
| 572 | |
---|
| 573 | |
---|
[684] | 574 | bmat local_max(0,0); |
---|
| 575 | int to, muto = 0; |
---|
| 576 | |
---|
[646] | 577 | //% statistics: |
---|
[684] | 578 | //double cputime0 = cputime; ?????????????????????????????????????????????????????????????? |
---|
[646] | 579 | //if nargout>=3; |
---|
[684] | 580 | |
---|
| 581 | CPU_Timer timer; |
---|
| 582 | timer.start(); |
---|
| 583 | |
---|
| 584 | ivec mutos(max_nrep+2); |
---|
| 585 | vec maxmutos(max_nrep+2); |
---|
| 586 | mutos.zeros(); |
---|
| 587 | maxmutos.zeros(); |
---|
[646] | 588 | |
---|
| 589 | |
---|
| 590 | //end; |
---|
| 591 | //% ---------------------- |
---|
[684] | 592 | |
---|
[646] | 593 | //% For stopping-rule calculation |
---|
| 594 | //%so = 2^(n_data -1-length(belief_in)- length(belief_out)); % do we use this ? |
---|
| 595 | //% ---------------------- |
---|
[684] | 596 | |
---|
| 597 | ivec all_str = linspace(1,n_data); |
---|
| 598 | |
---|
| 599 | Array<str_aux> global_best(1); |
---|
| 600 | global_best(0) = full; |
---|
[646] | 601 | |
---|
| 602 | |
---|
| 603 | //% MAIN LOOP is here. |
---|
| 604 | |
---|
[684] | 605 | str_aux best; |
---|
| 606 | for (int n_start = -1; n_start <= max_nrep; n_start++){ |
---|
| 607 | str_aux last,best; |
---|
| 608 | |
---|
| 609 | to = n_start+2; |
---|
| 610 | |
---|
| 611 | if (n_start == -1){ |
---|
| 612 | //% start from the full structure |
---|
| 613 | last = full; |
---|
| 614 | } |
---|
| 615 | else {if (n_start == 0) |
---|
| 616 | //% start from the empty structure |
---|
| 617 | last = empty; |
---|
| 618 | |
---|
| 619 | else{ |
---|
| 620 | //% start from random structure |
---|
| 621 | |
---|
| 622 | ivec last_str = find(to_bvec<int>(::concat<int>(0,floor_i(2*randu(n_data-1)))));// this creates random vector consisting of indexes, and sorted |
---|
| 623 | last = sestrremove(full,setdiff(all_str,::concat<int>(::concat<int>(1 ,last_str), empty.strRgr))); |
---|
| 624 | |
---|
| 625 | } |
---|
| 626 | } |
---|
| 627 | //% DEBUGging print: |
---|
| 628 | //%fprintf('STRUCTURE generated in loop %2i was %s\n', n_start, strPrintstr(last)); |
---|
| 629 | |
---|
| 630 | //% The loop is repeated until likelihood stops growing (break condition |
---|
| 631 | //% used at the end; |
---|
| 632 | |
---|
| 633 | |
---|
| 634 | while (1){ |
---|
| 635 | //% This structure is going to hold the best elements |
---|
| 636 | best = last; |
---|
| 637 | //% Nesting by removing elements (enpoorment) |
---|
| 638 | ivec removed_items = setdiff(last.strRgr,belief_in); |
---|
| 639 | |
---|
| 640 | ivec removed_item; |
---|
| 641 | str_aux newone; |
---|
| 642 | |
---|
| 643 | for (int i = 0; i < removed_items.length(); i++){ |
---|
| 644 | removed_item = vec_1(removed_items(i)); |
---|
| 645 | newone = sestrremove(last,removed_item);//????????????????????????????? |
---|
| 646 | if (nbest>1){ |
---|
| 647 | add_new(global_best,newone,nbest); |
---|
[646] | 648 | } |
---|
[684] | 649 | if (newone.loglik>best.loglik){ |
---|
| 650 | best = newone; |
---|
[646] | 651 | } |
---|
[684] | 652 | } |
---|
| 653 | //% Nesting by adding elements (enrichment) |
---|
| 654 | ivec added_items = setdiff(last.strMis,belief_out); |
---|
| 655 | ivec added_item; |
---|
| 656 | |
---|
| 657 | for (int j = 0; j < added_items.length(); j++){ |
---|
| 658 | added_item = vec_1(added_items(j)); |
---|
| 659 | newone = sestrinsert(last,added_item); |
---|
| 660 | if (nbest>1){ |
---|
| 661 | add_new(global_best,newone,nbest); |
---|
[646] | 662 | } |
---|
[684] | 663 | if (newone.loglik>best.loglik){ |
---|
| 664 | best = newone; |
---|
| 665 | } |
---|
| 666 | } |
---|
[646] | 667 | |
---|
[684] | 668 | |
---|
[646] | 669 | |
---|
[684] | 670 | |
---|
[646] | 671 | |
---|
[684] | 672 | //% Break condition if likelihood does not change. |
---|
| 673 | if (best.loglik <= last.loglik) |
---|
| 674 | break; |
---|
| 675 | else |
---|
| 676 | //% Making best structure last structure. |
---|
| 677 | last = best; |
---|
| 678 | |
---|
| 679 | |
---|
| 680 | } |
---|
| 681 | |
---|
[646] | 682 | |
---|
[684] | 683 | |
---|
[646] | 684 | |
---|
| 685 | |
---|
[684] | 686 | // % DEBUGging print: |
---|
| 687 | //%fprintf('STRUCTURE found (local maxima) in loop %2i was %s randun_seed=%11lu randun_counter=%4lu\n', n_start, strPrintstr(best), randn('seed'), RANDUN_COUNTER); |
---|
| 688 | |
---|
| 689 | //% Collecting of the best structure in case we don't need the second parameter |
---|
| 690 | if (nbest<=1){ |
---|
| 691 | if (best.loglik > global_best(0).loglik){ |
---|
| 692 | global_best = best; |
---|
| 693 | } |
---|
| 694 | } |
---|
| 695 | |
---|
| 696 | //% uniqueness of the structure found |
---|
| 697 | int append = 1; |
---|
| 698 | |
---|
[646] | 699 | |
---|
[684] | 700 | for(int j = 0; j < local_max.rows() ; j++){ |
---|
| 701 | if (best.bitstr == local_max.get_row(j)){ |
---|
| 702 | append = 0; |
---|
| 703 | break; |
---|
| 704 | } |
---|
| 705 | } |
---|
[646] | 706 | |
---|
[684] | 707 | |
---|
| 708 | if (append){ |
---|
| 709 | local_max.append_row(best.bitstr); |
---|
| 710 | muto = muto + 1; |
---|
| 711 | } |
---|
| 712 | |
---|
| 713 | //% stopping rule: |
---|
| 714 | double maxmuto = (to-order_k-1)/lambda-to+1; |
---|
| 715 | if (to>2){ |
---|
| 716 | if (maxmuto>=muto){ |
---|
| 717 | //% fprintf('*'); |
---|
| 718 | break; |
---|
| 719 | } |
---|
| 720 | } |
---|
| 721 | |
---|
| 722 | // do statistics if necessary: |
---|
| 723 | //if (nargout>=3){ |
---|
| 724 | mutos(to-1) = muto; |
---|
| 725 | maxmutos(to-1) = maxmuto; |
---|
| 726 | //} |
---|
| 727 | } |
---|
| 728 | |
---|
[646] | 729 | //% Aftermath: The best structure was in: global_best |
---|
[684] | 730 | |
---|
[646] | 731 | //% Updating loglikelihoods: we have to add the constant stuff |
---|
| 732 | |
---|
| 733 | |
---|
| 734 | |
---|
[684] | 735 | for (int f=0 ; f <global_best.length(); f++){ |
---|
| 736 | global_best(f).loglik = global_best(f).loglik + seloglik2(global_best(f)); |
---|
| 737 | } |
---|
| 738 | |
---|
| 739 | /*for f=1:length(global_best); |
---|
| 740 | global_best(f).loglik = global_best(f).loglik + seloglik2(global_best(f)); |
---|
| 741 | end;*/ |
---|
[646] | 742 | |
---|
| 743 | |
---|
| 744 | //% Making first output parameter: |
---|
| 745 | |
---|
[684] | 746 | int max_i = 0; |
---|
| 747 | for (int j = 1; j < global_best.length(); j++) |
---|
| 748 | if (global_best(max_i).loglik < (global_best(j).loglik)) max_i = j; |
---|
[646] | 749 | |
---|
[684] | 750 | best = global_best(max_i); |
---|
| 751 | cout << endl << endl << endl << endl << best.strRgr; |
---|
| 752 | cout << endl << endl << endl << endl << best.strRgr; |
---|
[646] | 753 | //% Making the second output parameter |
---|
| 754 | |
---|
[684] | 755 | vec logliks(global_best.length()); |
---|
| 756 | for (int j = 0; j < logliks.length(); j++) |
---|
| 757 | logliks(j) = global_best(j).loglik; |
---|
[646] | 758 | |
---|
[684] | 759 | ivec i = sort_index(logliks); |
---|
| 760 | rgrsout.set_length(global_best.length()); |
---|
[646] | 761 | |
---|
[684] | 762 | for (int j = global_best.length() - 1; j >= 0; j--) |
---|
| 763 | rgrsout(j) = global_best(i(j)); |
---|
| 764 | |
---|
[646] | 765 | //if (nargout>=3); |
---|
[684] | 766 | |
---|
| 767 | |
---|
| 768 | str_statistics statistics; |
---|
[646] | 769 | |
---|
[684] | 770 | statistics.allstrs = 2^(n_data -1-length(belief_in) - length(belief_out)); |
---|
| 771 | statistics.nrand = to-2; |
---|
| 772 | statistics.unique = muto; |
---|
| 773 | statistics.to = to; |
---|
| 774 | statistics.cputime_seconds = timer.get_time(); |
---|
| 775 | statistics.itemspeed = statistics.to / statistics.cputime_seconds; |
---|
| 776 | statistics.muto = muto; |
---|
| 777 | statistics.mutos = mutos; |
---|
| 778 | statistics.maxmutos = maxmutos; |
---|
[646] | 779 | //end; |
---|
| 780 | |
---|
[684] | 781 | return best.strRgr; |
---|
[646] | 782 | |
---|
[684] | 783 | } |
---|
[646] | 784 | |
---|
| 785 | #ifdef LADIM |
---|
| 786 | //% randun seed stuff: |
---|
| 787 | //%randn('seed',SEED); |
---|
[684] | 788 | |
---|
[646] | 789 | //% --------------------- END of MAIN program -------------------- |
---|
[684] | 790 | |
---|
[646] | 791 | % This is needed for bitstr manipulations |
---|
| 792 | /*function out = str_bitset(in,ns,nbits) |
---|
| 793 | out = in; |
---|
| 794 | for n = ns; |
---|
| 795 | index = 1+floor((n-2)/nbits); |
---|
| 796 | bitindex = 1+rem(n-2,nbits); |
---|
| 797 | out(index) = bitset(out(index),bitindex); |
---|
[684] | 798 | end; |
---|
[646] | 799 | function out = str_bitres(in,ns,nbits) |
---|
| 800 | out = in; |
---|
| 801 | for n = ns; |
---|
| 802 | index = 1+floor((n-2)/nbits); |
---|
| 803 | bitindex = 1+rem(n-2,nbits); |
---|
| 804 | mask = bitset(0,bitindex); |
---|
| 805 | out(index) = bitxor(bitor(out(index),mask),mask); |
---|
| 806 | end;*/ |
---|
[684] | 807 | |
---|
| 808 | function out = strPrintstr(in) |
---|
| 809 | out = '0'; |
---|
| 810 | nbits = in.nbits; |
---|
| 811 | for f = 2:length(in.d0); |
---|
| 812 | index = 1+floor((f-2)/nbits); |
---|
| 813 | bitindex = 1+rem(f-2,nbits); |
---|
| 814 | if bitget(in.bitstr(index),bitindex); |
---|
| 815 | out(f) = '1'; |
---|
| 816 | else; |
---|
| 817 | out(f) = '0'; |
---|
| 818 | end; |
---|
| 819 | end; |
---|
| 820 | |
---|
[646] | 821 | /*function global_best_out = add_new(global_best,new,nbest) |
---|
| 822 | % Eventually add to global best, but do not go over nbest values |
---|
| 823 | % Also avoids repeating things, which makes this function awfully slow |
---|
| 824 | if length(global_best)>=nbest; |
---|
| 825 | logliks = [global_best.loglik]; |
---|
| 826 | [loglik i] = min(logliks); |
---|
| 827 | global_best_out = global_best; |
---|
| 828 | if loglik<new.loglik; |
---|
| 829 | % if ~any(logliks == new.loglik); |
---|
| 830 | addit=1; |
---|
| 831 | for f = [global_best.bitstr]; |
---|
| 832 | if f == new.bitstr; |
---|
| 833 | addit = 0; |
---|
| 834 | break; |
---|
| 835 | end; |
---|
[684] | 836 | end; |
---|
[646] | 837 | if addit; |
---|
| 838 | global_best_out(i) = new; |
---|
| 839 | % DEBUGging print: |
---|
| 840 | % fprintf('ADDED structure, add_new: %s, loglik=%g\n', strPrintstr(new), new.loglik); |
---|
[684] | 841 | end; |
---|
[646] | 842 | end; |
---|
| 843 | else; |
---|
| 844 | global_best_out = [global_best new]; |
---|
| 845 | end;*/ |
---|
[684] | 846 | |
---|
[646] | 847 | /*function out = sestrremove(in,removed_elements); |
---|
| 848 | % Removes elements from regressor |
---|
| 849 | n_strL = length(in.strL); |
---|
| 850 | out = in; |
---|
| 851 | for f=removed_elements; |
---|
| 852 | posit1 = find(out.strL==1); |
---|
| 853 | positf = find(out.strL==f); |
---|
| 854 | for g=(positf-1):-1:posit1; |
---|
| 855 | % BEGIN: We are swapping g and g+1 NOW!!!! |
---|
| 856 | [out.L, out.d] = seswapudl(out.L, out.d, g); |
---|
| 857 | [out.L0, out.d0] = seswapudl(out.L0, out.d0, g); |
---|
| 858 | out.strL([g g+1]) = [out.strL(g+1) out.strL(g)]; |
---|
| 859 | % END |
---|
| 860 | end; |
---|
| 861 | end; |
---|
| 862 | out.posit1 = find(out.strL==1); |
---|
| 863 | out.strRgr = out.strL((out.posit1+1):n_strL); |
---|
| 864 | out.strMis = out.strL(1:(out.posit1-1)); |
---|
| 865 | out.bitstr = str_bitres(out.bitstr,removed_elements,out.nbits); |
---|
| 866 | out.loglik = seloglik1(out);*/ |
---|
[684] | 867 | |
---|
[646] | 868 | /*function out = sestrinsert(in,inserted_elements); |
---|
| 869 | % Moves elements into regressor |
---|
| 870 | n_strL = length(in.strL); |
---|
| 871 | out = in; |
---|
| 872 | for f=inserted_elements; |
---|
| 873 | posit1 = find(out.strL==1); |
---|
| 874 | positf = find(out.strL==f); |
---|
| 875 | for g=positf:(posit1-1); |
---|
| 876 | % BEGIN: We are swapping g and g+1 NOW!!!! |
---|
| 877 | [out.L, out.d] = seswapudl(out.L, out.d, g); |
---|
| 878 | [out.L0, out.d0] = seswapudl(out.L0, out.d0, g); |
---|
| 879 | out.strL([g g+1]) = [out.strL(g+1) out.strL(g)]; |
---|
| 880 | % END |
---|
| 881 | end; |
---|
[684] | 882 | end; |
---|
[646] | 883 | out.posit1 = find(out.strL==1); |
---|
| 884 | out.strRgr = out.strL((out.posit1+1):n_strL); |
---|
| 885 | out.strMis = out.strL(1:(out.posit1-1)); |
---|
| 886 | out.bitstr = str_bitset(out.bitstr,inserted_elements,out.nbits); |
---|
| 887 | out.loglik = seloglik1(out);*/ |
---|
[684] | 888 | |
---|
[646] | 889 | % |
---|
| 890 | % seloglik_real = seloglik1 + seloglik2 |
---|
[684] | 891 | % |
---|
| 892 | |
---|
| 893 | /*function l = seloglik1(in) |
---|
| 894 | % This is the loglikelihood (non-constant part) - this should be used in |
---|
| 895 | % frequent computation |
---|
| 896 | len = length(in.d); |
---|
| 897 | p1 = in.posit1; |
---|
| 898 | |
---|
| 899 | i1 = -0.5*in.nu *log(in.d (p1)) -0.5*sum(log(in.d ((p1+1):len))); |
---|
| 900 | i0 = -0.5*in.nu0*log(in.d0(p1)) -0.5*sum(log(in.d0((p1+1):len))); |
---|
| 901 | l = i1-i0; |
---|
| 902 | |
---|
| 903 | % DEBUGGing print: |
---|
| 904 | % fprintf('SELOGLIK1: str=%s loglik=%g\n', strPrintstr(in), l);*/ |
---|
| 905 | |
---|
| 906 | |
---|
| 907 | function l = seloglik2(in) |
---|
| 908 | % This is the loglikelihood (constant part) - this should be added to |
---|
| 909 | % everything at the end. It needs some computation, so it is useless to |
---|
| 910 | % make it for all the stuff |
---|
| 911 | logpi = log(pi); |
---|
| 912 | |
---|
| 913 | i1 = gammaln(in.nu /2) - 0.5*in.nu *logpi; |
---|
| 914 | i0 = gammaln(in.nu0/2) - 0.5*in.nu0*logpi; |
---|
| 915 | l = i1-i0; |
---|
| 916 | |
---|
| 917 | |
---|
[646] | 918 | /*function [Lout, dout] = seswapudl(L,d,i); |
---|
| 919 | %SESWAPUDL swaps information matrix in decomposition V=L^T diag(d) L |
---|
| 920 | % |
---|
| 921 | % [Lout, dout] = seswapudl(L,d,i); |
---|
| 922 | % |
---|
| 923 | % L : lower triangular matrix with 1's on diagonal of the decomposistion |
---|
| 924 | % d : diagonal vector of diagonal matrix of the decomposition |
---|
[684] | 925 | % i : index of line to be swapped with the next one |
---|
[646] | 926 | % Lout : output lower triangular matrix |
---|
| 927 | % dout : output diagional vector of diagonal matrix D |
---|
| 928 | % |
---|
| 929 | % Description: |
---|
| 930 | % Lout' * diag(dout) * Lout = P(i,i+1) * L' * diag(d) * L * P(i,i+1); |
---|
[684] | 931 | % |
---|
[646] | 932 | % Where permutation matrix P(i,j) permutates columns if applied from the |
---|
| 933 | % right and line if applied from the left. |
---|
[684] | 934 | % |
---|
[646] | 935 | % Note: naming: |
---|
| 936 | % se = structure estimation |
---|
| 937 | % lite = light, simple |
---|
| 938 | % udl = U*D*L, or more precisely, L'*D*L, also called as ld |
---|
[684] | 939 | % |
---|
[646] | 940 | % Design : L. Tesar |
---|
| 941 | % Updated : Feb 2003 |
---|
| 942 | % Project : post-ProDaCTool |
---|
| 943 | % Reference: sedydr |
---|
[684] | 944 | |
---|
[646] | 945 | j = i+1; |
---|
[684] | 946 | |
---|
[646] | 947 | pomd = d(i); |
---|
| 948 | d(i) = d(j); |
---|
| 949 | d(j) = pomd; |
---|
[684] | 950 | |
---|
[646] | 951 | pomL = L(i,:); |
---|
| 952 | L(i,:) = L(j,:); |
---|
| 953 | L(j,:) = pomL; |
---|
[684] | 954 | |
---|
[646] | 955 | pomL = L(:,i); |
---|
| 956 | L(:,i) = L(:,j); |
---|
| 957 | L(:,j) = pomL; |
---|
[684] | 958 | |
---|
[646] | 959 | % We must be working with LINES of matrix L ! |
---|
[684] | 960 | |
---|
| 961 | r = L(i,:)'; |
---|
[646] | 962 | f = L(j,:)'; |
---|
| 963 | Dr = d(i); |
---|
| 964 | Df = d(j); |
---|
[684] | 965 | |
---|
[646] | 966 | [r, f, Dr, Df] = sedydr(r, f, Dr, Df, j); |
---|
[684] | 967 | |
---|
[646] | 968 | r0 = r(i); |
---|
| 969 | Dr = Dr*r0*r0; |
---|
| 970 | r = r/r0; |
---|
[684] | 971 | |
---|
| 972 | L(i,:) = r'; |
---|
[646] | 973 | L(j,:) = f'; |
---|
| 974 | d(i) = Dr; |
---|
| 975 | d(j) = Df; |
---|
[684] | 976 | |
---|
[646] | 977 | L(i,i) = 1; |
---|
| 978 | L(j,j) = 1; |
---|
[684] | 979 | |
---|
[646] | 980 | Lout = L; |
---|
| 981 | dout = d;*/ |
---|
[684] | 982 | |
---|
[646] | 983 | /*function [rout, fout, Drout, Dfout, kr] = sedydr(r,f,Dr,Df,R,jl,jh); |
---|
| 984 | %SEDYDR dyadic reduction, performs transformation of sum of 2 dyads |
---|
| 985 | % |
---|
| 986 | % [rout, fout, Drout, Dfout, kr] = sedydr(r,f,Dr,Df,R,jl,jh); |
---|
| 987 | % [rout, fout, Drout, Dfout] = sedydr(r,f,Dr,Df,R); |
---|
| 988 | % |
---|
[684] | 989 | % Description: dyadic reduction, performs transformation of sum of |
---|
[646] | 990 | % 2 dyads r*Dr*r'+ f*Df*f' so that the element of r pointed by R is zeroed |
---|
| 991 | % |
---|
| 992 | % r : column vector of reduced dyad |
---|
| 993 | % f : column vector of reducing dyad |
---|
| 994 | % Dr : scalar with weight of reduced dyad |
---|
| 995 | % Df : scalar with weight of reducing dyad |
---|
| 996 | % R : scalar number giving 1 based index to the element of r, |
---|
[684] | 997 | % which is to be reduced to |
---|
[646] | 998 | % zero; the corresponding element of f is assumed to be 1. |
---|
[684] | 999 | % jl : lower index of the range within which the dyads are |
---|
[646] | 1000 | % modified (can be omitted, then everything is updated) |
---|
[684] | 1001 | % jh : upper index of the range within which the dyads are |
---|
[646] | 1002 | % modified (can be omitted then everything is updated) |
---|
| 1003 | % rout,fout,Drout,dfout : resulting two dyads |
---|
| 1004 | % kr : coefficient used in the transformation of r |
---|
| 1005 | % rnew = r + kr*f |
---|
| 1006 | % |
---|
[684] | 1007 | % Description: dyadic reduction, performs transformation of sum of |
---|
[646] | 1008 | % 2 dyads r*Dr*r'+ f*Df*f' so that the element of r indexed by R is zeroed |
---|
| 1009 | % Remark1: Constant mzero means machine zero and should be modified |
---|
| 1010 | % according to the precision of particular machine |
---|
| 1011 | % Remark2: jl and jh are, in fact, obsolete. It takes longer time to |
---|
| 1012 | % compute them compared to plain version. The reason is that we |
---|
| 1013 | % are doing vector operations in m-file. Other reason is that |
---|
| 1014 | % we need to copy whole vector anyway. It can save half of time for |
---|
| 1015 | % c-file, if you use it correctly. (please do tests) |
---|
| 1016 | % |
---|
| 1017 | % Note: naming: |
---|
| 1018 | % se = structure estimation |
---|
| 1019 | % dydr = dyadic reduction |
---|
| 1020 | % |
---|
| 1021 | % Original Fortran design: V. Peterka 17-7-89 |
---|
| 1022 | % Modified for c-language: probably R. Kulhavy |
---|
| 1023 | % Modified for m-language: L. Tesar 2/2003 |
---|
| 1024 | % Updated: Feb 2003 |
---|
| 1025 | % Project: post-ProDaCTool |
---|
| 1026 | % Reference: none |
---|
[684] | 1027 | |
---|
[646] | 1028 | if nargin<6; |
---|
| 1029 | update_whole=1; |
---|
| 1030 | else |
---|
| 1031 | update_whole=0; |
---|
| 1032 | end; |
---|
[684] | 1033 | |
---|
[646] | 1034 | mzero = 1e-32; |
---|
[684] | 1035 | |
---|
[646] | 1036 | if Dr<mzero; |
---|
| 1037 | Dr=0; |
---|
| 1038 | end; |
---|
[684] | 1039 | |
---|
[646] | 1040 | r0 = r(R); |
---|
| 1041 | kD = Df; |
---|
| 1042 | kr = r0 * Dr; |
---|
| 1043 | Dfout = kD + r0 * kr; |
---|
[684] | 1044 | |
---|
[646] | 1045 | if Dfout > mzero; |
---|
| 1046 | kD = kD / Dfout; |
---|
| 1047 | kr = kr / Dfout; |
---|
| 1048 | else; |
---|
| 1049 | kD = 1; |
---|
| 1050 | kr = 0; |
---|
| 1051 | end; |
---|
[684] | 1052 | |
---|
[646] | 1053 | Drout = Dr * kD; |
---|
[684] | 1054 | |
---|
[646] | 1055 | % Try to uncomment marked stuff (*) if in numerical problems, but I don't |
---|
| 1056 | % think it can make any difference for normal healthy floating-point unit |
---|
| 1057 | if update_whole; |
---|
| 1058 | rout = r - r0*f; |
---|
| 1059 | % rout(R) = 0; % * could be needed for some nonsense cases(or numeric reasons?), normally not |
---|
| 1060 | fout = f + kr*rout; |
---|
| 1061 | % fout(R) = 1; % * could be needed for some nonsense cases(or numeric reasons?), normally not |
---|
[684] | 1062 | else; |
---|
[646] | 1063 | rout = r; |
---|
| 1064 | fout = f; |
---|
| 1065 | rout(jl:jh) = r(jl:jh) - r0 * f(jl:jh); |
---|
[684] | 1066 | rout(R) = 0; |
---|
[646] | 1067 | fout(jl:jh) = f(jl:jh) + kr * rout(jl:jh); |
---|
| 1068 | end;*/ |
---|
[684] | 1069 | |
---|
| 1070 | |
---|
| 1071 | |
---|
[646] | 1072 | #endif |
---|
[684] | 1073 | |
---|
| 1074 | |
---|
[646] | 1075 | } |
---|