[3839] | 1 | /*************************************************************************
|
---|
| 2 | Copyright (c) 2007-2008, Sergey Bochkanov (ALGLIB project).
|
---|
| 3 |
|
---|
| 4 | >>> SOURCE LICENSE >>>
|
---|
| 5 | This program is free software; you can redistribute it and/or modify
|
---|
| 6 | it under the terms of the GNU General Public License as published by
|
---|
| 7 | the Free Software Foundation (www.fsf.org); either version 2 of the
|
---|
| 8 | License, or (at your option) any later version.
|
---|
| 9 |
|
---|
| 10 | This program is distributed in the hope that it will be useful,
|
---|
| 11 | but WITHOUT ANY WARRANTY; without even the implied warranty of
|
---|
| 12 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
---|
| 13 | GNU General Public License for more details.
|
---|
| 14 |
|
---|
| 15 | A copy of the GNU General Public License is available at
|
---|
| 16 | http://www.fsf.org/licensing/licenses
|
---|
| 17 |
|
---|
| 18 | >>> END OF LICENSE >>>
|
---|
| 19 | *************************************************************************/
|
---|
| 20 |
|
---|
| 21 | using System;
|
---|
| 22 |
|
---|
| 23 | namespace alglib
|
---|
| 24 | {
|
---|
| 25 | public class linreg
|
---|
| 26 | {
|
---|
| 27 | public struct linearmodel
|
---|
| 28 | {
|
---|
| 29 | public double[] w;
|
---|
| 30 | };
|
---|
| 31 |
|
---|
| 32 |
|
---|
| 33 | /*************************************************************************
|
---|
| 34 | LRReport structure contains additional information about linear model:
|
---|
| 35 | * C - covariation matrix, array[0..NVars,0..NVars].
|
---|
| 36 | C[i,j] = Cov(A[i],A[j])
|
---|
| 37 | * RMSError - root mean square error on a training set
|
---|
| 38 | * AvgError - average error on a training set
|
---|
| 39 | * AvgRelError - average relative error on a training set (excluding
|
---|
| 40 | observations with zero function value).
|
---|
| 41 | * CVRMSError - leave-one-out cross-validation estimate of
|
---|
| 42 | generalization error. Calculated using fast algorithm
|
---|
| 43 | with O(NVars*NPoints) complexity.
|
---|
| 44 | * CVAvgError - cross-validation estimate of average error
|
---|
| 45 | * CVAvgRelError - cross-validation estimate of average relative error
|
---|
| 46 |
|
---|
| 47 | All other fields of the structure are intended for internal use and should
|
---|
| 48 | not be used outside ALGLIB.
|
---|
| 49 | *************************************************************************/
|
---|
| 50 | public struct lrreport
|
---|
| 51 | {
|
---|
| 52 | public double[,] c;
|
---|
| 53 | public double rmserror;
|
---|
| 54 | public double avgerror;
|
---|
| 55 | public double avgrelerror;
|
---|
| 56 | public double cvrmserror;
|
---|
| 57 | public double cvavgerror;
|
---|
| 58 | public double cvavgrelerror;
|
---|
| 59 | public int ncvdefects;
|
---|
| 60 | public int[] cvdefects;
|
---|
| 61 | };
|
---|
| 62 |
|
---|
| 63 |
|
---|
| 64 |
|
---|
| 65 |
|
---|
| 66 | public const int lrvnum = 5;
|
---|
| 67 |
|
---|
| 68 |
|
---|
| 69 | /*************************************************************************
|
---|
| 70 | Linear regression
|
---|
| 71 |
|
---|
| 72 | Subroutine builds model:
|
---|
| 73 |
|
---|
| 74 | Y = A(0)*X[0] + ... + A(N-1)*X[N-1] + A(N)
|
---|
| 75 |
|
---|
| 76 | and model found in ALGLIB format, covariation matrix, training set errors
|
---|
| 77 | (rms, average, average relative) and leave-one-out cross-validation
|
---|
| 78 | estimate of the generalization error. CV estimate calculated using fast
|
---|
| 79 | algorithm with O(NPoints*NVars) complexity.
|
---|
| 80 |
|
---|
| 81 | When covariation matrix is calculated standard deviations of function
|
---|
| 82 | values are assumed to be equal to RMS error on the training set.
|
---|
| 83 |
|
---|
| 84 | INPUT PARAMETERS:
|
---|
| 85 | XY - training set, array [0..NPoints-1,0..NVars]:
|
---|
| 86 | * NVars columns - independent variables
|
---|
| 87 | * last column - dependent variable
|
---|
| 88 | NPoints - training set size, NPoints>NVars+1
|
---|
| 89 | NVars - number of independent variables
|
---|
| 90 |
|
---|
| 91 | OUTPUT PARAMETERS:
|
---|
| 92 | Info - return code:
|
---|
| 93 | * -255, in case of unknown internal error
|
---|
| 94 | * -4, if internal SVD subroutine haven't converged
|
---|
| 95 | * -1, if incorrect parameters was passed (NPoints<NVars+2, NVars<1).
|
---|
| 96 | * 1, if subroutine successfully finished
|
---|
| 97 | LM - linear model in the ALGLIB format. Use subroutines of
|
---|
| 98 | this unit to work with the model.
|
---|
| 99 | AR - additional results
|
---|
| 100 |
|
---|
| 101 |
|
---|
| 102 | -- ALGLIB --
|
---|
| 103 | Copyright 02.08.2008 by Bochkanov Sergey
|
---|
| 104 | *************************************************************************/
|
---|
| 105 | public static void lrbuild(ref double[,] xy,
|
---|
| 106 | int npoints,
|
---|
| 107 | int nvars,
|
---|
| 108 | ref int info,
|
---|
| 109 | ref linearmodel lm,
|
---|
| 110 | ref lrreport ar)
|
---|
| 111 | {
|
---|
| 112 | double[] s = new double[0];
|
---|
| 113 | int i = 0;
|
---|
| 114 | double sigma2 = 0;
|
---|
| 115 | int i_ = 0;
|
---|
| 116 |
|
---|
| 117 | if( npoints<=nvars+1 | nvars<1 )
|
---|
| 118 | {
|
---|
| 119 | info = -1;
|
---|
| 120 | return;
|
---|
| 121 | }
|
---|
| 122 | s = new double[npoints-1+1];
|
---|
| 123 | for(i=0; i<=npoints-1; i++)
|
---|
| 124 | {
|
---|
| 125 | s[i] = 1;
|
---|
| 126 | }
|
---|
| 127 | lrbuilds(ref xy, ref s, npoints, nvars, ref info, ref lm, ref ar);
|
---|
| 128 | if( info<0 )
|
---|
| 129 | {
|
---|
| 130 | return;
|
---|
| 131 | }
|
---|
| 132 | sigma2 = AP.Math.Sqr(ar.rmserror)*npoints/(npoints-nvars-1);
|
---|
| 133 | for(i=0; i<=nvars; i++)
|
---|
| 134 | {
|
---|
| 135 | for(i_=0; i_<=nvars;i_++)
|
---|
| 136 | {
|
---|
| 137 | ar.c[i,i_] = sigma2*ar.c[i,i_];
|
---|
| 138 | }
|
---|
| 139 | }
|
---|
| 140 | }
|
---|
| 141 |
|
---|
| 142 |
|
---|
| 143 | /*************************************************************************
|
---|
| 144 | Linear regression
|
---|
| 145 |
|
---|
| 146 | Variant of LRBuild which uses vector of standatd deviations (errors in
|
---|
| 147 | function values).
|
---|
| 148 |
|
---|
| 149 | INPUT PARAMETERS:
|
---|
| 150 | XY - training set, array [0..NPoints-1,0..NVars]:
|
---|
| 151 | * NVars columns - independent variables
|
---|
| 152 | * last column - dependent variable
|
---|
| 153 | S - standard deviations (errors in function values)
|
---|
| 154 | array[0..NPoints-1], S[i]>0.
|
---|
| 155 | NPoints - training set size, NPoints>NVars+1
|
---|
| 156 | NVars - number of independent variables
|
---|
| 157 |
|
---|
| 158 | OUTPUT PARAMETERS:
|
---|
| 159 | Info - return code:
|
---|
| 160 | * -255, in case of unknown internal error
|
---|
| 161 | * -4, if internal SVD subroutine haven't converged
|
---|
| 162 | * -1, if incorrect parameters was passed (NPoints<NVars+2, NVars<1).
|
---|
| 163 | * -2, if S[I]<=0
|
---|
| 164 | * 1, if subroutine successfully finished
|
---|
| 165 | LM - linear model in the ALGLIB format. Use subroutines of
|
---|
| 166 | this unit to work with the model.
|
---|
| 167 | AR - additional results
|
---|
| 168 |
|
---|
| 169 |
|
---|
| 170 | -- ALGLIB --
|
---|
| 171 | Copyright 02.08.2008 by Bochkanov Sergey
|
---|
| 172 | *************************************************************************/
|
---|
| 173 | public static void lrbuilds(ref double[,] xy,
|
---|
| 174 | ref double[] s,
|
---|
| 175 | int npoints,
|
---|
| 176 | int nvars,
|
---|
| 177 | ref int info,
|
---|
| 178 | ref linearmodel lm,
|
---|
| 179 | ref lrreport ar)
|
---|
| 180 | {
|
---|
| 181 | double[,] xyi = new double[0,0];
|
---|
| 182 | double[] x = new double[0];
|
---|
| 183 | double[] means = new double[0];
|
---|
| 184 | double[] sigmas = new double[0];
|
---|
| 185 | int i = 0;
|
---|
| 186 | int j = 0;
|
---|
| 187 | double v = 0;
|
---|
| 188 | int offs = 0;
|
---|
| 189 | double mean = 0;
|
---|
| 190 | double variance = 0;
|
---|
| 191 | double skewness = 0;
|
---|
| 192 | double kurtosis = 0;
|
---|
| 193 | int i_ = 0;
|
---|
| 194 |
|
---|
| 195 |
|
---|
| 196 | //
|
---|
| 197 | // Test parameters
|
---|
| 198 | //
|
---|
| 199 | if( npoints<=nvars+1 | nvars<1 )
|
---|
| 200 | {
|
---|
| 201 | info = -1;
|
---|
| 202 | return;
|
---|
| 203 | }
|
---|
| 204 |
|
---|
| 205 | //
|
---|
| 206 | // Copy data, add one more column (constant term)
|
---|
| 207 | //
|
---|
| 208 | xyi = new double[npoints-1+1, nvars+1+1];
|
---|
| 209 | for(i=0; i<=npoints-1; i++)
|
---|
| 210 | {
|
---|
| 211 | for(i_=0; i_<=nvars-1;i_++)
|
---|
| 212 | {
|
---|
| 213 | xyi[i,i_] = xy[i,i_];
|
---|
| 214 | }
|
---|
| 215 | xyi[i,nvars] = 1;
|
---|
| 216 | xyi[i,nvars+1] = xy[i,nvars];
|
---|
| 217 | }
|
---|
| 218 |
|
---|
| 219 | //
|
---|
| 220 | // Standartization
|
---|
| 221 | //
|
---|
| 222 | x = new double[npoints-1+1];
|
---|
| 223 | means = new double[nvars-1+1];
|
---|
| 224 | sigmas = new double[nvars-1+1];
|
---|
| 225 | for(j=0; j<=nvars-1; j++)
|
---|
| 226 | {
|
---|
| 227 | for(i_=0; i_<=npoints-1;i_++)
|
---|
| 228 | {
|
---|
| 229 | x[i_] = xy[i_,j];
|
---|
| 230 | }
|
---|
| 231 | descriptivestatistics.calculatemoments(ref x, npoints, ref mean, ref variance, ref skewness, ref kurtosis);
|
---|
| 232 | means[j] = mean;
|
---|
| 233 | sigmas[j] = Math.Sqrt(variance);
|
---|
| 234 | if( (double)(sigmas[j])==(double)(0) )
|
---|
| 235 | {
|
---|
| 236 | sigmas[j] = 1;
|
---|
| 237 | }
|
---|
| 238 | for(i=0; i<=npoints-1; i++)
|
---|
| 239 | {
|
---|
| 240 | xyi[i,j] = (xyi[i,j]-means[j])/sigmas[j];
|
---|
| 241 | }
|
---|
| 242 | }
|
---|
| 243 |
|
---|
| 244 | //
|
---|
| 245 | // Internal processing
|
---|
| 246 | //
|
---|
| 247 | lrinternal(ref xyi, ref s, npoints, nvars+1, ref info, ref lm, ref ar);
|
---|
| 248 | if( info<0 )
|
---|
| 249 | {
|
---|
| 250 | return;
|
---|
| 251 | }
|
---|
| 252 |
|
---|
| 253 | //
|
---|
| 254 | // Un-standartization
|
---|
| 255 | //
|
---|
| 256 | offs = (int)Math.Round(lm.w[3]);
|
---|
| 257 | for(j=0; j<=nvars-1; j++)
|
---|
| 258 | {
|
---|
| 259 |
|
---|
| 260 | //
|
---|
| 261 | // Constant term is updated (and its covariance too,
|
---|
| 262 | // since it gets some variance from J-th component)
|
---|
| 263 | //
|
---|
| 264 | lm.w[offs+nvars] = lm.w[offs+nvars]-lm.w[offs+j]*means[j]/sigmas[j];
|
---|
| 265 | v = means[j]/sigmas[j];
|
---|
| 266 | for(i_=0; i_<=nvars;i_++)
|
---|
| 267 | {
|
---|
| 268 | ar.c[nvars,i_] = ar.c[nvars,i_] - v*ar.c[j,i_];
|
---|
| 269 | }
|
---|
| 270 | for(i_=0; i_<=nvars;i_++)
|
---|
| 271 | {
|
---|
| 272 | ar.c[i_,nvars] = ar.c[i_,nvars] - v*ar.c[i_,j];
|
---|
| 273 | }
|
---|
| 274 |
|
---|
| 275 | //
|
---|
| 276 | // J-th term is updated
|
---|
| 277 | //
|
---|
| 278 | lm.w[offs+j] = lm.w[offs+j]/sigmas[j];
|
---|
| 279 | v = 1/sigmas[j];
|
---|
| 280 | for(i_=0; i_<=nvars;i_++)
|
---|
| 281 | {
|
---|
| 282 | ar.c[j,i_] = v*ar.c[j,i_];
|
---|
| 283 | }
|
---|
| 284 | for(i_=0; i_<=nvars;i_++)
|
---|
| 285 | {
|
---|
| 286 | ar.c[i_,j] = v*ar.c[i_,j];
|
---|
| 287 | }
|
---|
| 288 | }
|
---|
| 289 | }
|
---|
| 290 |
|
---|
| 291 |
|
---|
| 292 | /*************************************************************************
|
---|
| 293 | Like LRBuildS, but builds model
|
---|
| 294 |
|
---|
| 295 | Y = A(0)*X[0] + ... + A(N-1)*X[N-1]
|
---|
| 296 |
|
---|
| 297 | i.e. with zero constant term.
|
---|
| 298 |
|
---|
| 299 | -- ALGLIB --
|
---|
| 300 | Copyright 30.10.2008 by Bochkanov Sergey
|
---|
| 301 | *************************************************************************/
|
---|
| 302 | public static void lrbuildzs(ref double[,] xy,
|
---|
| 303 | ref double[] s,
|
---|
| 304 | int npoints,
|
---|
| 305 | int nvars,
|
---|
| 306 | ref int info,
|
---|
| 307 | ref linearmodel lm,
|
---|
| 308 | ref lrreport ar)
|
---|
| 309 | {
|
---|
| 310 | double[,] xyi = new double[0,0];
|
---|
| 311 | double[] x = new double[0];
|
---|
| 312 | double[] c = new double[0];
|
---|
| 313 | int i = 0;
|
---|
| 314 | int j = 0;
|
---|
| 315 | double v = 0;
|
---|
| 316 | int offs = 0;
|
---|
| 317 | double mean = 0;
|
---|
| 318 | double variance = 0;
|
---|
| 319 | double skewness = 0;
|
---|
| 320 | double kurtosis = 0;
|
---|
| 321 | int i_ = 0;
|
---|
| 322 |
|
---|
| 323 |
|
---|
| 324 | //
|
---|
| 325 | // Test parameters
|
---|
| 326 | //
|
---|
| 327 | if( npoints<=nvars+1 | nvars<1 )
|
---|
| 328 | {
|
---|
| 329 | info = -1;
|
---|
| 330 | return;
|
---|
| 331 | }
|
---|
| 332 |
|
---|
| 333 | //
|
---|
| 334 | // Copy data, add one more column (constant term)
|
---|
| 335 | //
|
---|
| 336 | xyi = new double[npoints-1+1, nvars+1+1];
|
---|
| 337 | for(i=0; i<=npoints-1; i++)
|
---|
| 338 | {
|
---|
| 339 | for(i_=0; i_<=nvars-1;i_++)
|
---|
| 340 | {
|
---|
| 341 | xyi[i,i_] = xy[i,i_];
|
---|
| 342 | }
|
---|
| 343 | xyi[i,nvars] = 0;
|
---|
| 344 | xyi[i,nvars+1] = xy[i,nvars];
|
---|
| 345 | }
|
---|
| 346 |
|
---|
| 347 | //
|
---|
| 348 | // Standartization: unusual scaling
|
---|
| 349 | //
|
---|
| 350 | x = new double[npoints-1+1];
|
---|
| 351 | c = new double[nvars-1+1];
|
---|
| 352 | for(j=0; j<=nvars-1; j++)
|
---|
| 353 | {
|
---|
| 354 | for(i_=0; i_<=npoints-1;i_++)
|
---|
| 355 | {
|
---|
| 356 | x[i_] = xy[i_,j];
|
---|
| 357 | }
|
---|
| 358 | descriptivestatistics.calculatemoments(ref x, npoints, ref mean, ref variance, ref skewness, ref kurtosis);
|
---|
| 359 | if( (double)(Math.Abs(mean))>(double)(Math.Sqrt(variance)) )
|
---|
| 360 | {
|
---|
| 361 |
|
---|
| 362 | //
|
---|
| 363 | // variation is relatively small, it is better to
|
---|
| 364 | // bring mean value to 1
|
---|
| 365 | //
|
---|
| 366 | c[j] = mean;
|
---|
| 367 | }
|
---|
| 368 | else
|
---|
| 369 | {
|
---|
| 370 |
|
---|
| 371 | //
|
---|
| 372 | // variation is large, it is better to bring variance to 1
|
---|
| 373 | //
|
---|
| 374 | if( (double)(variance)==(double)(0) )
|
---|
| 375 | {
|
---|
| 376 | variance = 1;
|
---|
| 377 | }
|
---|
| 378 | c[j] = Math.Sqrt(variance);
|
---|
| 379 | }
|
---|
| 380 | for(i=0; i<=npoints-1; i++)
|
---|
| 381 | {
|
---|
| 382 | xyi[i,j] = xyi[i,j]/c[j];
|
---|
| 383 | }
|
---|
| 384 | }
|
---|
| 385 |
|
---|
| 386 | //
|
---|
| 387 | // Internal processing
|
---|
| 388 | //
|
---|
| 389 | lrinternal(ref xyi, ref s, npoints, nvars+1, ref info, ref lm, ref ar);
|
---|
| 390 | if( info<0 )
|
---|
| 391 | {
|
---|
| 392 | return;
|
---|
| 393 | }
|
---|
| 394 |
|
---|
| 395 | //
|
---|
| 396 | // Un-standartization
|
---|
| 397 | //
|
---|
| 398 | offs = (int)Math.Round(lm.w[3]);
|
---|
| 399 | for(j=0; j<=nvars-1; j++)
|
---|
| 400 | {
|
---|
| 401 |
|
---|
| 402 | //
|
---|
| 403 | // J-th term is updated
|
---|
| 404 | //
|
---|
| 405 | lm.w[offs+j] = lm.w[offs+j]/c[j];
|
---|
| 406 | v = 1/c[j];
|
---|
| 407 | for(i_=0; i_<=nvars;i_++)
|
---|
| 408 | {
|
---|
| 409 | ar.c[j,i_] = v*ar.c[j,i_];
|
---|
| 410 | }
|
---|
| 411 | for(i_=0; i_<=nvars;i_++)
|
---|
| 412 | {
|
---|
| 413 | ar.c[i_,j] = v*ar.c[i_,j];
|
---|
| 414 | }
|
---|
| 415 | }
|
---|
| 416 | }
|
---|
| 417 |
|
---|
| 418 |
|
---|
| 419 | /*************************************************************************
|
---|
| 420 | Like LRBuild but builds model
|
---|
| 421 |
|
---|
| 422 | Y = A(0)*X[0] + ... + A(N-1)*X[N-1]
|
---|
| 423 |
|
---|
| 424 | i.e. with zero constant term.
|
---|
| 425 |
|
---|
| 426 | -- ALGLIB --
|
---|
| 427 | Copyright 30.10.2008 by Bochkanov Sergey
|
---|
| 428 | *************************************************************************/
|
---|
| 429 | public static void lrbuildz(ref double[,] xy,
|
---|
| 430 | int npoints,
|
---|
| 431 | int nvars,
|
---|
| 432 | ref int info,
|
---|
| 433 | ref linearmodel lm,
|
---|
| 434 | ref lrreport ar)
|
---|
| 435 | {
|
---|
| 436 | double[] s = new double[0];
|
---|
| 437 | int i = 0;
|
---|
| 438 | double sigma2 = 0;
|
---|
| 439 | int i_ = 0;
|
---|
| 440 |
|
---|
| 441 | if( npoints<=nvars+1 | nvars<1 )
|
---|
| 442 | {
|
---|
| 443 | info = -1;
|
---|
| 444 | return;
|
---|
| 445 | }
|
---|
| 446 | s = new double[npoints-1+1];
|
---|
| 447 | for(i=0; i<=npoints-1; i++)
|
---|
| 448 | {
|
---|
| 449 | s[i] = 1;
|
---|
| 450 | }
|
---|
| 451 | lrbuildzs(ref xy, ref s, npoints, nvars, ref info, ref lm, ref ar);
|
---|
| 452 | if( info<0 )
|
---|
| 453 | {
|
---|
| 454 | return;
|
---|
| 455 | }
|
---|
| 456 | sigma2 = AP.Math.Sqr(ar.rmserror)*npoints/(npoints-nvars-1);
|
---|
| 457 | for(i=0; i<=nvars; i++)
|
---|
| 458 | {
|
---|
| 459 | for(i_=0; i_<=nvars;i_++)
|
---|
| 460 | {
|
---|
| 461 | ar.c[i,i_] = sigma2*ar.c[i,i_];
|
---|
| 462 | }
|
---|
| 463 | }
|
---|
| 464 | }
|
---|
| 465 |
|
---|
| 466 |
|
---|
| 467 | /*************************************************************************
|
---|
| 468 | Unpacks coefficients of linear model.
|
---|
| 469 |
|
---|
| 470 | INPUT PARAMETERS:
|
---|
| 471 | LM - linear model in ALGLIB format
|
---|
| 472 |
|
---|
| 473 | OUTPUT PARAMETERS:
|
---|
| 474 | V - coefficients, array[0..NVars]
|
---|
| 475 | NVars - number of independent variables (one less than number
|
---|
| 476 | of coefficients)
|
---|
| 477 |
|
---|
| 478 | -- ALGLIB --
|
---|
| 479 | Copyright 30.08.2008 by Bochkanov Sergey
|
---|
| 480 | *************************************************************************/
|
---|
| 481 | public static void lrunpack(ref linearmodel lm,
|
---|
| 482 | ref double[] v,
|
---|
| 483 | ref int nvars)
|
---|
| 484 | {
|
---|
| 485 | int offs = 0;
|
---|
| 486 | int i_ = 0;
|
---|
| 487 | int i1_ = 0;
|
---|
| 488 |
|
---|
| 489 | System.Diagnostics.Debug.Assert((int)Math.Round(lm.w[1])==lrvnum, "LINREG: Incorrect LINREG version!");
|
---|
| 490 | nvars = (int)Math.Round(lm.w[2]);
|
---|
| 491 | offs = (int)Math.Round(lm.w[3]);
|
---|
| 492 | v = new double[nvars+1];
|
---|
| 493 | i1_ = (offs) - (0);
|
---|
| 494 | for(i_=0; i_<=nvars;i_++)
|
---|
| 495 | {
|
---|
| 496 | v[i_] = lm.w[i_+i1_];
|
---|
| 497 | }
|
---|
| 498 | }
|
---|
| 499 |
|
---|
| 500 |
|
---|
| 501 | /*************************************************************************
|
---|
| 502 | "Packs" coefficients and creates linear model in ALGLIB format (LRUnpack
|
---|
| 503 | reversed).
|
---|
| 504 |
|
---|
| 505 | INPUT PARAMETERS:
|
---|
| 506 | V - coefficients, array[0..NVars]
|
---|
| 507 | NVars - number of independent variables
|
---|
| 508 |
|
---|
| 509 | OUTPUT PAREMETERS:
|
---|
| 510 | LM - linear model.
|
---|
| 511 |
|
---|
| 512 | -- ALGLIB --
|
---|
| 513 | Copyright 30.08.2008 by Bochkanov Sergey
|
---|
| 514 | *************************************************************************/
|
---|
| 515 | public static void lrpack(ref double[] v,
|
---|
| 516 | int nvars,
|
---|
| 517 | ref linearmodel lm)
|
---|
| 518 | {
|
---|
| 519 | int offs = 0;
|
---|
| 520 | int i_ = 0;
|
---|
| 521 | int i1_ = 0;
|
---|
| 522 |
|
---|
| 523 | lm.w = new double[4+nvars+1];
|
---|
| 524 | offs = 4;
|
---|
| 525 | lm.w[0] = 4+nvars+1;
|
---|
| 526 | lm.w[1] = lrvnum;
|
---|
| 527 | lm.w[2] = nvars;
|
---|
| 528 | lm.w[3] = offs;
|
---|
| 529 | i1_ = (0) - (offs);
|
---|
| 530 | for(i_=offs; i_<=offs+nvars;i_++)
|
---|
| 531 | {
|
---|
| 532 | lm.w[i_] = v[i_+i1_];
|
---|
| 533 | }
|
---|
| 534 | }
|
---|
| 535 |
|
---|
| 536 |
|
---|
| 537 | /*************************************************************************
|
---|
| 538 | Procesing
|
---|
| 539 |
|
---|
| 540 | INPUT PARAMETERS:
|
---|
| 541 | LM - linear model
|
---|
| 542 | X - input vector, array[0..NVars-1].
|
---|
| 543 |
|
---|
| 544 | Result:
|
---|
| 545 | value of linear model regression estimate
|
---|
| 546 |
|
---|
| 547 | -- ALGLIB --
|
---|
| 548 | Copyright 03.09.2008 by Bochkanov Sergey
|
---|
| 549 | *************************************************************************/
|
---|
| 550 | public static double lrprocess(ref linearmodel lm,
|
---|
| 551 | ref double[] x)
|
---|
| 552 | {
|
---|
| 553 | double result = 0;
|
---|
| 554 | double v = 0;
|
---|
| 555 | int offs = 0;
|
---|
| 556 | int nvars = 0;
|
---|
| 557 | int i_ = 0;
|
---|
| 558 | int i1_ = 0;
|
---|
| 559 |
|
---|
| 560 | System.Diagnostics.Debug.Assert((int)Math.Round(lm.w[1])==lrvnum, "LINREG: Incorrect LINREG version!");
|
---|
| 561 | nvars = (int)Math.Round(lm.w[2]);
|
---|
| 562 | offs = (int)Math.Round(lm.w[3]);
|
---|
| 563 | i1_ = (offs)-(0);
|
---|
| 564 | v = 0.0;
|
---|
| 565 | for(i_=0; i_<=nvars-1;i_++)
|
---|
| 566 | {
|
---|
| 567 | v += x[i_]*lm.w[i_+i1_];
|
---|
| 568 | }
|
---|
| 569 | result = v+lm.w[offs+nvars];
|
---|
| 570 | return result;
|
---|
| 571 | }
|
---|
| 572 |
|
---|
| 573 |
|
---|
| 574 | /*************************************************************************
|
---|
| 575 | RMS error on the test set
|
---|
| 576 |
|
---|
| 577 | INPUT PARAMETERS:
|
---|
| 578 | LM - linear model
|
---|
| 579 | XY - test set
|
---|
| 580 | NPoints - test set size
|
---|
| 581 |
|
---|
| 582 | RESULT:
|
---|
| 583 | root mean square error.
|
---|
| 584 |
|
---|
| 585 | -- ALGLIB --
|
---|
| 586 | Copyright 30.08.2008 by Bochkanov Sergey
|
---|
| 587 | *************************************************************************/
|
---|
| 588 | public static double lrrmserror(ref linearmodel lm,
|
---|
| 589 | ref double[,] xy,
|
---|
| 590 | int npoints)
|
---|
| 591 | {
|
---|
| 592 | double result = 0;
|
---|
| 593 | int i = 0;
|
---|
| 594 | double v = 0;
|
---|
| 595 | int offs = 0;
|
---|
| 596 | int nvars = 0;
|
---|
| 597 | int i_ = 0;
|
---|
| 598 | int i1_ = 0;
|
---|
| 599 |
|
---|
| 600 | System.Diagnostics.Debug.Assert((int)Math.Round(lm.w[1])==lrvnum, "LINREG: Incorrect LINREG version!");
|
---|
| 601 | nvars = (int)Math.Round(lm.w[2]);
|
---|
| 602 | offs = (int)Math.Round(lm.w[3]);
|
---|
| 603 | result = 0;
|
---|
| 604 | for(i=0; i<=npoints-1; i++)
|
---|
| 605 | {
|
---|
| 606 | i1_ = (offs)-(0);
|
---|
| 607 | v = 0.0;
|
---|
| 608 | for(i_=0; i_<=nvars-1;i_++)
|
---|
| 609 | {
|
---|
| 610 | v += xy[i,i_]*lm.w[i_+i1_];
|
---|
| 611 | }
|
---|
| 612 | v = v+lm.w[offs+nvars];
|
---|
| 613 | result = result+AP.Math.Sqr(v-xy[i,nvars]);
|
---|
| 614 | }
|
---|
| 615 | result = Math.Sqrt(result/npoints);
|
---|
| 616 | return result;
|
---|
| 617 | }
|
---|
| 618 |
|
---|
| 619 |
|
---|
| 620 | /*************************************************************************
|
---|
| 621 | Average error on the test set
|
---|
| 622 |
|
---|
| 623 | INPUT PARAMETERS:
|
---|
| 624 | LM - linear model
|
---|
| 625 | XY - test set
|
---|
| 626 | NPoints - test set size
|
---|
| 627 |
|
---|
| 628 | RESULT:
|
---|
| 629 | average error.
|
---|
| 630 |
|
---|
| 631 | -- ALGLIB --
|
---|
| 632 | Copyright 30.08.2008 by Bochkanov Sergey
|
---|
| 633 | *************************************************************************/
|
---|
| 634 | public static double lravgerror(ref linearmodel lm,
|
---|
| 635 | ref double[,] xy,
|
---|
| 636 | int npoints)
|
---|
| 637 | {
|
---|
| 638 | double result = 0;
|
---|
| 639 | int i = 0;
|
---|
| 640 | double v = 0;
|
---|
| 641 | int offs = 0;
|
---|
| 642 | int nvars = 0;
|
---|
| 643 | int i_ = 0;
|
---|
| 644 | int i1_ = 0;
|
---|
| 645 |
|
---|
| 646 | System.Diagnostics.Debug.Assert((int)Math.Round(lm.w[1])==lrvnum, "LINREG: Incorrect LINREG version!");
|
---|
| 647 | nvars = (int)Math.Round(lm.w[2]);
|
---|
| 648 | offs = (int)Math.Round(lm.w[3]);
|
---|
| 649 | result = 0;
|
---|
| 650 | for(i=0; i<=npoints-1; i++)
|
---|
| 651 | {
|
---|
| 652 | i1_ = (offs)-(0);
|
---|
| 653 | v = 0.0;
|
---|
| 654 | for(i_=0; i_<=nvars-1;i_++)
|
---|
| 655 | {
|
---|
| 656 | v += xy[i,i_]*lm.w[i_+i1_];
|
---|
| 657 | }
|
---|
| 658 | v = v+lm.w[offs+nvars];
|
---|
| 659 | result = result+Math.Abs(v-xy[i,nvars]);
|
---|
| 660 | }
|
---|
| 661 | result = result/npoints;
|
---|
| 662 | return result;
|
---|
| 663 | }
|
---|
| 664 |
|
---|
| 665 |
|
---|
| 666 | /*************************************************************************
|
---|
| 667 | RMS error on the test set
|
---|
| 668 |
|
---|
| 669 | INPUT PARAMETERS:
|
---|
| 670 | LM - linear model
|
---|
| 671 | XY - test set
|
---|
| 672 | NPoints - test set size
|
---|
| 673 |
|
---|
| 674 | RESULT:
|
---|
| 675 | average relative error.
|
---|
| 676 |
|
---|
| 677 | -- ALGLIB --
|
---|
| 678 | Copyright 30.08.2008 by Bochkanov Sergey
|
---|
| 679 | *************************************************************************/
|
---|
| 680 | public static double lravgrelerror(ref linearmodel lm,
|
---|
| 681 | ref double[,] xy,
|
---|
| 682 | int npoints)
|
---|
| 683 | {
|
---|
| 684 | double result = 0;
|
---|
| 685 | int i = 0;
|
---|
| 686 | int k = 0;
|
---|
| 687 | double v = 0;
|
---|
| 688 | int offs = 0;
|
---|
| 689 | int nvars = 0;
|
---|
| 690 | int i_ = 0;
|
---|
| 691 | int i1_ = 0;
|
---|
| 692 |
|
---|
| 693 | System.Diagnostics.Debug.Assert((int)Math.Round(lm.w[1])==lrvnum, "LINREG: Incorrect LINREG version!");
|
---|
| 694 | nvars = (int)Math.Round(lm.w[2]);
|
---|
| 695 | offs = (int)Math.Round(lm.w[3]);
|
---|
| 696 | result = 0;
|
---|
| 697 | k = 0;
|
---|
| 698 | for(i=0; i<=npoints-1; i++)
|
---|
| 699 | {
|
---|
| 700 | if( (double)(xy[i,nvars])!=(double)(0) )
|
---|
| 701 | {
|
---|
| 702 | i1_ = (offs)-(0);
|
---|
| 703 | v = 0.0;
|
---|
| 704 | for(i_=0; i_<=nvars-1;i_++)
|
---|
| 705 | {
|
---|
| 706 | v += xy[i,i_]*lm.w[i_+i1_];
|
---|
| 707 | }
|
---|
| 708 | v = v+lm.w[offs+nvars];
|
---|
| 709 | result = result+Math.Abs((v-xy[i,nvars])/xy[i,nvars]);
|
---|
| 710 | k = k+1;
|
---|
| 711 | }
|
---|
| 712 | }
|
---|
| 713 | if( k!=0 )
|
---|
| 714 | {
|
---|
| 715 | result = result/k;
|
---|
| 716 | }
|
---|
| 717 | return result;
|
---|
| 718 | }
|
---|
| 719 |
|
---|
| 720 |
|
---|
| 721 | /*************************************************************************
|
---|
| 722 | Copying of LinearModel strucure
|
---|
| 723 |
|
---|
| 724 | INPUT PARAMETERS:
|
---|
| 725 | LM1 - original
|
---|
| 726 |
|
---|
| 727 | OUTPUT PARAMETERS:
|
---|
| 728 | LM2 - copy
|
---|
| 729 |
|
---|
| 730 | -- ALGLIB --
|
---|
| 731 | Copyright 15.03.2009 by Bochkanov Sergey
|
---|
| 732 | *************************************************************************/
|
---|
| 733 | public static void lrcopy(ref linearmodel lm1,
|
---|
| 734 | ref linearmodel lm2)
|
---|
| 735 | {
|
---|
| 736 | int k = 0;
|
---|
| 737 | int i_ = 0;
|
---|
| 738 |
|
---|
| 739 | k = (int)Math.Round(lm1.w[0]);
|
---|
| 740 | lm2.w = new double[k-1+1];
|
---|
| 741 | for(i_=0; i_<=k-1;i_++)
|
---|
| 742 | {
|
---|
| 743 | lm2.w[i_] = lm1.w[i_];
|
---|
| 744 | }
|
---|
| 745 | }
|
---|
| 746 |
|
---|
| 747 |
|
---|
| 748 | /*************************************************************************
|
---|
| 749 | Serialization of LinearModel strucure
|
---|
| 750 |
|
---|
| 751 | INPUT PARAMETERS:
|
---|
| 752 | LM - original
|
---|
| 753 |
|
---|
| 754 | OUTPUT PARAMETERS:
|
---|
| 755 | RA - array of real numbers which stores model,
|
---|
| 756 | array[0..RLen-1]
|
---|
| 757 | RLen - RA lenght
|
---|
| 758 |
|
---|
| 759 | -- ALGLIB --
|
---|
| 760 | Copyright 15.03.2009 by Bochkanov Sergey
|
---|
| 761 | *************************************************************************/
|
---|
| 762 | public static void lrserialize(ref linearmodel lm,
|
---|
| 763 | ref double[] ra,
|
---|
| 764 | ref int rlen)
|
---|
| 765 | {
|
---|
| 766 | int i_ = 0;
|
---|
| 767 | int i1_ = 0;
|
---|
| 768 |
|
---|
| 769 | rlen = (int)Math.Round(lm.w[0])+1;
|
---|
| 770 | ra = new double[rlen-1+1];
|
---|
| 771 | ra[0] = lrvnum;
|
---|
| 772 | i1_ = (0) - (1);
|
---|
| 773 | for(i_=1; i_<=rlen-1;i_++)
|
---|
| 774 | {
|
---|
| 775 | ra[i_] = lm.w[i_+i1_];
|
---|
| 776 | }
|
---|
| 777 | }
|
---|
| 778 |
|
---|
| 779 |
|
---|
| 780 | /*************************************************************************
|
---|
| 781 | Unserialization of DecisionForest strucure
|
---|
| 782 |
|
---|
| 783 | INPUT PARAMETERS:
|
---|
| 784 | RA - real array which stores decision forest
|
---|
| 785 |
|
---|
| 786 | OUTPUT PARAMETERS:
|
---|
| 787 | LM - unserialized structure
|
---|
| 788 |
|
---|
| 789 | -- ALGLIB --
|
---|
| 790 | Copyright 15.03.2009 by Bochkanov Sergey
|
---|
| 791 | *************************************************************************/
|
---|
| 792 | public static void lrunserialize(ref double[] ra,
|
---|
| 793 | ref linearmodel lm)
|
---|
| 794 | {
|
---|
| 795 | int i_ = 0;
|
---|
| 796 | int i1_ = 0;
|
---|
| 797 |
|
---|
| 798 | System.Diagnostics.Debug.Assert((int)Math.Round(ra[0])==lrvnum, "LRUnserialize: incorrect array!");
|
---|
| 799 | lm.w = new double[(int)Math.Round(ra[1])-1+1];
|
---|
| 800 | i1_ = (1) - (0);
|
---|
| 801 | for(i_=0; i_<=(int)Math.Round(ra[1])-1;i_++)
|
---|
| 802 | {
|
---|
| 803 | lm.w[i_] = ra[i_+i1_];
|
---|
| 804 | }
|
---|
| 805 | }
|
---|
| 806 |
|
---|
| 807 |
|
---|
| 808 | public static void lrlines(ref double[,] xy,
|
---|
| 809 | ref double[] s,
|
---|
| 810 | int n,
|
---|
| 811 | ref int info,
|
---|
| 812 | ref double a,
|
---|
| 813 | ref double b,
|
---|
| 814 | ref double vara,
|
---|
| 815 | ref double varb,
|
---|
| 816 | ref double covab,
|
---|
| 817 | ref double corrab,
|
---|
| 818 | ref double p)
|
---|
| 819 | {
|
---|
| 820 | int i = 0;
|
---|
| 821 | double ss = 0;
|
---|
| 822 | double sx = 0;
|
---|
| 823 | double sxx = 0;
|
---|
| 824 | double sy = 0;
|
---|
| 825 | double stt = 0;
|
---|
| 826 | double e1 = 0;
|
---|
| 827 | double e2 = 0;
|
---|
| 828 | double t = 0;
|
---|
| 829 | double chi2 = 0;
|
---|
| 830 |
|
---|
| 831 | if( n<2 )
|
---|
| 832 | {
|
---|
| 833 | info = -1;
|
---|
| 834 | return;
|
---|
| 835 | }
|
---|
| 836 | for(i=0; i<=n-1; i++)
|
---|
| 837 | {
|
---|
| 838 | if( (double)(s[i])<=(double)(0) )
|
---|
| 839 | {
|
---|
| 840 | info = -2;
|
---|
| 841 | return;
|
---|
| 842 | }
|
---|
| 843 | }
|
---|
| 844 | info = 1;
|
---|
| 845 |
|
---|
| 846 | //
|
---|
| 847 | // Calculate S, SX, SY, SXX
|
---|
| 848 | //
|
---|
| 849 | ss = 0;
|
---|
| 850 | sx = 0;
|
---|
| 851 | sy = 0;
|
---|
| 852 | sxx = 0;
|
---|
| 853 | for(i=0; i<=n-1; i++)
|
---|
| 854 | {
|
---|
| 855 | t = AP.Math.Sqr(s[i]);
|
---|
| 856 | ss = ss+1/t;
|
---|
| 857 | sx = sx+xy[i,0]/t;
|
---|
| 858 | sy = sy+xy[i,1]/t;
|
---|
| 859 | sxx = sxx+AP.Math.Sqr(xy[i,0])/t;
|
---|
| 860 | }
|
---|
| 861 |
|
---|
| 862 | //
|
---|
| 863 | // Test for condition number
|
---|
| 864 | //
|
---|
| 865 | t = Math.Sqrt(4*AP.Math.Sqr(sx)+AP.Math.Sqr(ss-sxx));
|
---|
| 866 | e1 = 0.5*(ss+sxx+t);
|
---|
| 867 | e2 = 0.5*(ss+sxx-t);
|
---|
| 868 | if( (double)(Math.Min(e1, e2))<=(double)(1000*AP.Math.MachineEpsilon*Math.Max(e1, e2)) )
|
---|
| 869 | {
|
---|
| 870 | info = -3;
|
---|
| 871 | return;
|
---|
| 872 | }
|
---|
| 873 |
|
---|
| 874 | //
|
---|
| 875 | // Calculate A, B
|
---|
| 876 | //
|
---|
| 877 | a = 0;
|
---|
| 878 | b = 0;
|
---|
| 879 | stt = 0;
|
---|
| 880 | for(i=0; i<=n-1; i++)
|
---|
| 881 | {
|
---|
| 882 | t = (xy[i,0]-sx/ss)/s[i];
|
---|
| 883 | b = b+t*xy[i,1]/s[i];
|
---|
| 884 | stt = stt+AP.Math.Sqr(t);
|
---|
| 885 | }
|
---|
| 886 | b = b/stt;
|
---|
| 887 | a = (sy-sx*b)/ss;
|
---|
| 888 |
|
---|
| 889 | //
|
---|
| 890 | // Calculate goodness-of-fit
|
---|
| 891 | //
|
---|
| 892 | if( n>2 )
|
---|
| 893 | {
|
---|
| 894 | chi2 = 0;
|
---|
| 895 | for(i=0; i<=n-1; i++)
|
---|
| 896 | {
|
---|
| 897 | chi2 = chi2+AP.Math.Sqr((xy[i,1]-a-b*xy[i,0])/s[i]);
|
---|
| 898 | }
|
---|
| 899 | p = igammaf.incompletegammac(((double)(n-2))/(double)(2), chi2/2);
|
---|
| 900 | }
|
---|
| 901 | else
|
---|
| 902 | {
|
---|
| 903 | p = 1;
|
---|
| 904 | }
|
---|
| 905 |
|
---|
| 906 | //
|
---|
| 907 | // Calculate other parameters
|
---|
| 908 | //
|
---|
| 909 | vara = (1+AP.Math.Sqr(sx)/(ss*stt))/ss;
|
---|
| 910 | varb = 1/stt;
|
---|
| 911 | covab = -(sx/(ss*stt));
|
---|
| 912 | corrab = covab/Math.Sqrt(vara*varb);
|
---|
| 913 | }
|
---|
| 914 |
|
---|
| 915 |
|
---|
| 916 | public static void lrline(ref double[,] xy,
|
---|
| 917 | int n,
|
---|
| 918 | ref int info,
|
---|
| 919 | ref double a,
|
---|
| 920 | ref double b)
|
---|
| 921 | {
|
---|
| 922 | double[] s = new double[0];
|
---|
| 923 | int i = 0;
|
---|
| 924 | double vara = 0;
|
---|
| 925 | double varb = 0;
|
---|
| 926 | double covab = 0;
|
---|
| 927 | double corrab = 0;
|
---|
| 928 | double p = 0;
|
---|
| 929 |
|
---|
| 930 | if( n<2 )
|
---|
| 931 | {
|
---|
| 932 | info = -1;
|
---|
| 933 | return;
|
---|
| 934 | }
|
---|
| 935 | s = new double[n-1+1];
|
---|
| 936 | for(i=0; i<=n-1; i++)
|
---|
| 937 | {
|
---|
| 938 | s[i] = 1;
|
---|
| 939 | }
|
---|
| 940 | lrlines(ref xy, ref s, n, ref info, ref a, ref b, ref vara, ref varb, ref covab, ref corrab, ref p);
|
---|
| 941 | }
|
---|
| 942 |
|
---|
| 943 |
|
---|
| 944 | /*************************************************************************
|
---|
| 945 | Internal linear regression subroutine
|
---|
| 946 | *************************************************************************/
|
---|
| 947 | private static void lrinternal(ref double[,] xy,
|
---|
| 948 | ref double[] s,
|
---|
| 949 | int npoints,
|
---|
| 950 | int nvars,
|
---|
| 951 | ref int info,
|
---|
| 952 | ref linearmodel lm,
|
---|
| 953 | ref lrreport ar)
|
---|
| 954 | {
|
---|
| 955 | double[,] a = new double[0,0];
|
---|
| 956 | double[,] u = new double[0,0];
|
---|
| 957 | double[,] vt = new double[0,0];
|
---|
| 958 | double[,] vm = new double[0,0];
|
---|
| 959 | double[,] xym = new double[0,0];
|
---|
| 960 | double[] b = new double[0];
|
---|
| 961 | double[] sv = new double[0];
|
---|
| 962 | double[] t = new double[0];
|
---|
| 963 | double[] svi = new double[0];
|
---|
| 964 | double[] work = new double[0];
|
---|
| 965 | int i = 0;
|
---|
| 966 | int j = 0;
|
---|
| 967 | int k = 0;
|
---|
| 968 | int ncv = 0;
|
---|
| 969 | int na = 0;
|
---|
| 970 | int nacv = 0;
|
---|
| 971 | double r = 0;
|
---|
| 972 | double p = 0;
|
---|
| 973 | double epstol = 0;
|
---|
| 974 | lrreport ar2 = new lrreport();
|
---|
| 975 | int offs = 0;
|
---|
| 976 | linearmodel tlm = new linearmodel();
|
---|
| 977 | int i_ = 0;
|
---|
| 978 | int i1_ = 0;
|
---|
| 979 |
|
---|
| 980 | epstol = 1000;
|
---|
| 981 |
|
---|
| 982 | //
|
---|
| 983 | // Check for errors in data
|
---|
| 984 | //
|
---|
| 985 | if( npoints<nvars | nvars<1 )
|
---|
| 986 | {
|
---|
| 987 | info = -1;
|
---|
| 988 | return;
|
---|
| 989 | }
|
---|
| 990 | for(i=0; i<=npoints-1; i++)
|
---|
| 991 | {
|
---|
| 992 | if( (double)(s[i])<=(double)(0) )
|
---|
| 993 | {
|
---|
| 994 | info = -2;
|
---|
| 995 | return;
|
---|
| 996 | }
|
---|
| 997 | }
|
---|
| 998 | info = 1;
|
---|
| 999 |
|
---|
| 1000 | //
|
---|
| 1001 | // Create design matrix
|
---|
| 1002 | //
|
---|
| 1003 | a = new double[npoints-1+1, nvars-1+1];
|
---|
| 1004 | b = new double[npoints-1+1];
|
---|
| 1005 | for(i=0; i<=npoints-1; i++)
|
---|
| 1006 | {
|
---|
| 1007 | r = 1/s[i];
|
---|
| 1008 | for(i_=0; i_<=nvars-1;i_++)
|
---|
| 1009 | {
|
---|
| 1010 | a[i,i_] = r*xy[i,i_];
|
---|
| 1011 | }
|
---|
| 1012 | b[i] = xy[i,nvars]/s[i];
|
---|
| 1013 | }
|
---|
| 1014 |
|
---|
| 1015 | //
|
---|
| 1016 | // Allocate W:
|
---|
| 1017 | // W[0] array size
|
---|
| 1018 | // W[1] version number, 0
|
---|
| 1019 | // W[2] NVars (minus 1, to be compatible with external representation)
|
---|
| 1020 | // W[3] coefficients offset
|
---|
| 1021 | //
|
---|
| 1022 | lm.w = new double[4+nvars-1+1];
|
---|
| 1023 | offs = 4;
|
---|
| 1024 | lm.w[0] = 4+nvars;
|
---|
| 1025 | lm.w[1] = lrvnum;
|
---|
| 1026 | lm.w[2] = nvars-1;
|
---|
| 1027 | lm.w[3] = offs;
|
---|
| 1028 |
|
---|
| 1029 | //
|
---|
| 1030 | // Solve problem using SVD:
|
---|
| 1031 | //
|
---|
| 1032 | // 0. check for degeneracy (different types)
|
---|
| 1033 | // 1. A = U*diag(sv)*V'
|
---|
| 1034 | // 2. T = b'*U
|
---|
| 1035 | // 3. w = SUM((T[i]/sv[i])*V[..,i])
|
---|
| 1036 | // 4. cov(wi,wj) = SUM(Vji*Vjk/sv[i]^2,K=1..M)
|
---|
| 1037 | //
|
---|
| 1038 | // see $15.4 of "Numerical Recipes in C" for more information
|
---|
| 1039 | //
|
---|
| 1040 | t = new double[nvars-1+1];
|
---|
| 1041 | svi = new double[nvars-1+1];
|
---|
| 1042 | ar.c = new double[nvars-1+1, nvars-1+1];
|
---|
| 1043 | vm = new double[nvars-1+1, nvars-1+1];
|
---|
| 1044 | if( !svd.rmatrixsvd(a, npoints, nvars, 1, 1, 2, ref sv, ref u, ref vt) )
|
---|
| 1045 | {
|
---|
| 1046 | info = -4;
|
---|
| 1047 | return;
|
---|
| 1048 | }
|
---|
| 1049 | if( (double)(sv[0])<=(double)(0) )
|
---|
| 1050 | {
|
---|
| 1051 |
|
---|
| 1052 | //
|
---|
| 1053 | // Degenerate case: zero design matrix.
|
---|
| 1054 | //
|
---|
| 1055 | for(i=offs; i<=offs+nvars-1; i++)
|
---|
| 1056 | {
|
---|
| 1057 | lm.w[i] = 0;
|
---|
| 1058 | }
|
---|
| 1059 | ar.rmserror = lrrmserror(ref lm, ref xy, npoints);
|
---|
| 1060 | ar.avgerror = lravgerror(ref lm, ref xy, npoints);
|
---|
| 1061 | ar.avgrelerror = lravgrelerror(ref lm, ref xy, npoints);
|
---|
| 1062 | ar.cvrmserror = ar.rmserror;
|
---|
| 1063 | ar.cvavgerror = ar.avgerror;
|
---|
| 1064 | ar.cvavgrelerror = ar.avgrelerror;
|
---|
| 1065 | ar.ncvdefects = 0;
|
---|
| 1066 | ar.cvdefects = new int[nvars-1+1];
|
---|
| 1067 | ar.c = new double[nvars-1+1, nvars-1+1];
|
---|
| 1068 | for(i=0; i<=nvars-1; i++)
|
---|
| 1069 | {
|
---|
| 1070 | for(j=0; j<=nvars-1; j++)
|
---|
| 1071 | {
|
---|
| 1072 | ar.c[i,j] = 0;
|
---|
| 1073 | }
|
---|
| 1074 | }
|
---|
| 1075 | return;
|
---|
| 1076 | }
|
---|
| 1077 | if( (double)(sv[nvars-1])<=(double)(epstol*AP.Math.MachineEpsilon*sv[0]) )
|
---|
| 1078 | {
|
---|
| 1079 |
|
---|
| 1080 | //
|
---|
| 1081 | // Degenerate case, non-zero design matrix.
|
---|
| 1082 | //
|
---|
| 1083 | // We can leave it and solve task in SVD least squares fashion.
|
---|
| 1084 | // Solution and covariance matrix will be obtained correctly,
|
---|
| 1085 | // but CV error estimates - will not. It is better to reduce
|
---|
| 1086 | // it to non-degenerate task and to obtain correct CV estimates.
|
---|
| 1087 | //
|
---|
| 1088 | for(k=nvars; k>=1; k--)
|
---|
| 1089 | {
|
---|
| 1090 | if( (double)(sv[k-1])>(double)(epstol*AP.Math.MachineEpsilon*sv[0]) )
|
---|
| 1091 | {
|
---|
| 1092 |
|
---|
| 1093 | //
|
---|
| 1094 | // Reduce
|
---|
| 1095 | //
|
---|
| 1096 | xym = new double[npoints-1+1, k+1];
|
---|
| 1097 | for(i=0; i<=npoints-1; i++)
|
---|
| 1098 | {
|
---|
| 1099 | for(j=0; j<=k-1; j++)
|
---|
| 1100 | {
|
---|
| 1101 | r = 0.0;
|
---|
| 1102 | for(i_=0; i_<=nvars-1;i_++)
|
---|
| 1103 | {
|
---|
| 1104 | r += xy[i,i_]*vt[j,i_];
|
---|
| 1105 | }
|
---|
| 1106 | xym[i,j] = r;
|
---|
| 1107 | }
|
---|
| 1108 | xym[i,k] = xy[i,nvars];
|
---|
| 1109 | }
|
---|
| 1110 |
|
---|
| 1111 | //
|
---|
| 1112 | // Solve
|
---|
| 1113 | //
|
---|
| 1114 | lrinternal(ref xym, ref s, npoints, k, ref info, ref tlm, ref ar2);
|
---|
| 1115 | if( info!=1 )
|
---|
| 1116 | {
|
---|
| 1117 | return;
|
---|
| 1118 | }
|
---|
| 1119 |
|
---|
| 1120 | //
|
---|
| 1121 | // Convert back to un-reduced format
|
---|
| 1122 | //
|
---|
| 1123 | for(j=0; j<=nvars-1; j++)
|
---|
| 1124 | {
|
---|
| 1125 | lm.w[offs+j] = 0;
|
---|
| 1126 | }
|
---|
| 1127 | for(j=0; j<=k-1; j++)
|
---|
| 1128 | {
|
---|
| 1129 | r = tlm.w[offs+j];
|
---|
| 1130 | i1_ = (0) - (offs);
|
---|
| 1131 | for(i_=offs; i_<=offs+nvars-1;i_++)
|
---|
| 1132 | {
|
---|
| 1133 | lm.w[i_] = lm.w[i_] + r*vt[j,i_+i1_];
|
---|
| 1134 | }
|
---|
| 1135 | }
|
---|
| 1136 | ar.rmserror = ar2.rmserror;
|
---|
| 1137 | ar.avgerror = ar2.avgerror;
|
---|
| 1138 | ar.avgrelerror = ar2.avgrelerror;
|
---|
| 1139 | ar.cvrmserror = ar2.cvrmserror;
|
---|
| 1140 | ar.cvavgerror = ar2.cvavgerror;
|
---|
| 1141 | ar.cvavgrelerror = ar2.cvavgrelerror;
|
---|
| 1142 | ar.ncvdefects = ar2.ncvdefects;
|
---|
| 1143 | ar.cvdefects = new int[nvars-1+1];
|
---|
| 1144 | for(j=0; j<=ar.ncvdefects-1; j++)
|
---|
| 1145 | {
|
---|
| 1146 | ar.cvdefects[j] = ar2.cvdefects[j];
|
---|
| 1147 | }
|
---|
| 1148 | ar.c = new double[nvars-1+1, nvars-1+1];
|
---|
| 1149 | work = new double[nvars+1];
|
---|
| 1150 | blas.matrixmatrixmultiply(ref ar2.c, 0, k-1, 0, k-1, false, ref vt, 0, k-1, 0, nvars-1, false, 1.0, ref vm, 0, k-1, 0, nvars-1, 0.0, ref work);
|
---|
| 1151 | blas.matrixmatrixmultiply(ref vt, 0, k-1, 0, nvars-1, true, ref vm, 0, k-1, 0, nvars-1, false, 1.0, ref ar.c, 0, nvars-1, 0, nvars-1, 0.0, ref work);
|
---|
| 1152 | return;
|
---|
| 1153 | }
|
---|
| 1154 | }
|
---|
| 1155 | info = -255;
|
---|
| 1156 | return;
|
---|
| 1157 | }
|
---|
| 1158 | for(i=0; i<=nvars-1; i++)
|
---|
| 1159 | {
|
---|
| 1160 | if( (double)(sv[i])>(double)(epstol*AP.Math.MachineEpsilon*sv[0]) )
|
---|
| 1161 | {
|
---|
| 1162 | svi[i] = 1/sv[i];
|
---|
| 1163 | }
|
---|
| 1164 | else
|
---|
| 1165 | {
|
---|
| 1166 | svi[i] = 0;
|
---|
| 1167 | }
|
---|
| 1168 | }
|
---|
| 1169 | for(i=0; i<=nvars-1; i++)
|
---|
| 1170 | {
|
---|
| 1171 | t[i] = 0;
|
---|
| 1172 | }
|
---|
| 1173 | for(i=0; i<=npoints-1; i++)
|
---|
| 1174 | {
|
---|
| 1175 | r = b[i];
|
---|
| 1176 | for(i_=0; i_<=nvars-1;i_++)
|
---|
| 1177 | {
|
---|
| 1178 | t[i_] = t[i_] + r*u[i,i_];
|
---|
| 1179 | }
|
---|
| 1180 | }
|
---|
| 1181 | for(i=0; i<=nvars-1; i++)
|
---|
| 1182 | {
|
---|
| 1183 | lm.w[offs+i] = 0;
|
---|
| 1184 | }
|
---|
| 1185 | for(i=0; i<=nvars-1; i++)
|
---|
| 1186 | {
|
---|
| 1187 | r = t[i]*svi[i];
|
---|
| 1188 | i1_ = (0) - (offs);
|
---|
| 1189 | for(i_=offs; i_<=offs+nvars-1;i_++)
|
---|
| 1190 | {
|
---|
| 1191 | lm.w[i_] = lm.w[i_] + r*vt[i,i_+i1_];
|
---|
| 1192 | }
|
---|
| 1193 | }
|
---|
| 1194 | for(j=0; j<=nvars-1; j++)
|
---|
| 1195 | {
|
---|
| 1196 | r = svi[j];
|
---|
| 1197 | for(i_=0; i_<=nvars-1;i_++)
|
---|
| 1198 | {
|
---|
| 1199 | vm[i_,j] = r*vt[j,i_];
|
---|
| 1200 | }
|
---|
| 1201 | }
|
---|
| 1202 | for(i=0; i<=nvars-1; i++)
|
---|
| 1203 | {
|
---|
| 1204 | for(j=i; j<=nvars-1; j++)
|
---|
| 1205 | {
|
---|
| 1206 | r = 0.0;
|
---|
| 1207 | for(i_=0; i_<=nvars-1;i_++)
|
---|
| 1208 | {
|
---|
| 1209 | r += vm[i,i_]*vm[j,i_];
|
---|
| 1210 | }
|
---|
| 1211 | ar.c[i,j] = r;
|
---|
| 1212 | ar.c[j,i] = r;
|
---|
| 1213 | }
|
---|
| 1214 | }
|
---|
| 1215 |
|
---|
| 1216 | //
|
---|
| 1217 | // Leave-1-out cross-validation error.
|
---|
| 1218 | //
|
---|
| 1219 | // NOTATIONS:
|
---|
| 1220 | // A design matrix
|
---|
| 1221 | // A*x = b original linear least squares task
|
---|
| 1222 | // U*S*V' SVD of A
|
---|
| 1223 | // ai i-th row of the A
|
---|
| 1224 | // bi i-th element of the b
|
---|
| 1225 | // xf solution of the original LLS task
|
---|
| 1226 | //
|
---|
| 1227 | // Cross-validation error of i-th element from a sample is
|
---|
| 1228 | // calculated using following formula:
|
---|
| 1229 | //
|
---|
| 1230 | // ERRi = ai*xf - (ai*xf-bi*(ui*ui'))/(1-ui*ui') (1)
|
---|
| 1231 | //
|
---|
| 1232 | // This formula can be derived from normal equations of the
|
---|
| 1233 | // original task
|
---|
| 1234 | //
|
---|
| 1235 | // (A'*A)x = A'*b (2)
|
---|
| 1236 | //
|
---|
| 1237 | // by applying modification (zeroing out i-th row of A) to (2):
|
---|
| 1238 | //
|
---|
| 1239 | // (A-ai)'*(A-ai) = (A-ai)'*b
|
---|
| 1240 | //
|
---|
| 1241 | // and using Sherman-Morrison formula for updating matrix inverse
|
---|
| 1242 | //
|
---|
| 1243 | // NOTE 1: b is not zeroed out since it is much simpler and
|
---|
| 1244 | // does not influence final result.
|
---|
| 1245 | //
|
---|
| 1246 | // NOTE 2: some design matrices A have such ui that 1-ui*ui'=0.
|
---|
| 1247 | // Formula (1) can't be applied for such cases and they are skipped
|
---|
| 1248 | // from CV calculation (which distorts resulting CV estimate).
|
---|
| 1249 | // But from the properties of U we can conclude that there can
|
---|
| 1250 | // be no more than NVars such vectors. Usually
|
---|
| 1251 | // NVars << NPoints, so in a normal case it only slightly
|
---|
| 1252 | // influences result.
|
---|
| 1253 | //
|
---|
| 1254 | ncv = 0;
|
---|
| 1255 | na = 0;
|
---|
| 1256 | nacv = 0;
|
---|
| 1257 | ar.rmserror = 0;
|
---|
| 1258 | ar.avgerror = 0;
|
---|
| 1259 | ar.avgrelerror = 0;
|
---|
| 1260 | ar.cvrmserror = 0;
|
---|
| 1261 | ar.cvavgerror = 0;
|
---|
| 1262 | ar.cvavgrelerror = 0;
|
---|
| 1263 | ar.ncvdefects = 0;
|
---|
| 1264 | ar.cvdefects = new int[nvars-1+1];
|
---|
| 1265 | for(i=0; i<=npoints-1; i++)
|
---|
| 1266 | {
|
---|
| 1267 |
|
---|
| 1268 | //
|
---|
| 1269 | // Error on a training set
|
---|
| 1270 | //
|
---|
| 1271 | i1_ = (offs)-(0);
|
---|
| 1272 | r = 0.0;
|
---|
| 1273 | for(i_=0; i_<=nvars-1;i_++)
|
---|
| 1274 | {
|
---|
| 1275 | r += xy[i,i_]*lm.w[i_+i1_];
|
---|
| 1276 | }
|
---|
| 1277 | ar.rmserror = ar.rmserror+AP.Math.Sqr(r-xy[i,nvars]);
|
---|
| 1278 | ar.avgerror = ar.avgerror+Math.Abs(r-xy[i,nvars]);
|
---|
| 1279 | if( (double)(xy[i,nvars])!=(double)(0) )
|
---|
| 1280 | {
|
---|
| 1281 | ar.avgrelerror = ar.avgrelerror+Math.Abs((r-xy[i,nvars])/xy[i,nvars]);
|
---|
| 1282 | na = na+1;
|
---|
| 1283 | }
|
---|
| 1284 |
|
---|
| 1285 | //
|
---|
| 1286 | // Error using fast leave-one-out cross-validation
|
---|
| 1287 | //
|
---|
| 1288 | p = 0.0;
|
---|
| 1289 | for(i_=0; i_<=nvars-1;i_++)
|
---|
| 1290 | {
|
---|
| 1291 | p += u[i,i_]*u[i,i_];
|
---|
| 1292 | }
|
---|
| 1293 | if( (double)(p)>(double)(1-epstol*AP.Math.MachineEpsilon) )
|
---|
| 1294 | {
|
---|
| 1295 | ar.cvdefects[ar.ncvdefects] = i;
|
---|
| 1296 | ar.ncvdefects = ar.ncvdefects+1;
|
---|
| 1297 | continue;
|
---|
| 1298 | }
|
---|
| 1299 | r = s[i]*(r/s[i]-b[i]*p)/(1-p);
|
---|
| 1300 | ar.cvrmserror = ar.cvrmserror+AP.Math.Sqr(r-xy[i,nvars]);
|
---|
| 1301 | ar.cvavgerror = ar.cvavgerror+Math.Abs(r-xy[i,nvars]);
|
---|
| 1302 | if( (double)(xy[i,nvars])!=(double)(0) )
|
---|
| 1303 | {
|
---|
| 1304 | ar.cvavgrelerror = ar.cvavgrelerror+Math.Abs((r-xy[i,nvars])/xy[i,nvars]);
|
---|
| 1305 | nacv = nacv+1;
|
---|
| 1306 | }
|
---|
| 1307 | ncv = ncv+1;
|
---|
| 1308 | }
|
---|
| 1309 | if( ncv==0 )
|
---|
| 1310 | {
|
---|
| 1311 |
|
---|
| 1312 | //
|
---|
| 1313 | // Something strange: ALL ui are degenerate.
|
---|
| 1314 | // Unexpected...
|
---|
| 1315 | //
|
---|
| 1316 | info = -255;
|
---|
| 1317 | return;
|
---|
| 1318 | }
|
---|
| 1319 | ar.rmserror = Math.Sqrt(ar.rmserror/npoints);
|
---|
| 1320 | ar.avgerror = ar.avgerror/npoints;
|
---|
| 1321 | if( na!=0 )
|
---|
| 1322 | {
|
---|
| 1323 | ar.avgrelerror = ar.avgrelerror/na;
|
---|
| 1324 | }
|
---|
| 1325 | ar.cvrmserror = Math.Sqrt(ar.cvrmserror/ncv);
|
---|
| 1326 | ar.cvavgerror = ar.cvavgerror/ncv;
|
---|
| 1327 | if( nacv!=0 )
|
---|
| 1328 | {
|
---|
| 1329 | ar.cvavgrelerror = ar.cvavgrelerror/nacv;
|
---|
| 1330 | }
|
---|
| 1331 | }
|
---|
| 1332 | }
|
---|
| 1333 | }
|
---|