[14991] | 1 | #region License Information
|
---|
| 2 | /* HeuristicLab
|
---|
| 3 | * Copyright (C) 2002-2016 Heuristic and Evolutionary Algorithms Laboratory (HEAL)
|
---|
| 4 | *
|
---|
| 5 | * This file is part of HeuristicLab.
|
---|
| 6 | *
|
---|
| 7 | * HeuristicLab is free software: you can redistribute it and/or modify
|
---|
| 8 | * it under the terms of the GNU General Public License as published by
|
---|
| 9 | * the Free Software Foundation, either version 3 of the License, or
|
---|
| 10 | * (at your option) any later version.
|
---|
| 11 | *
|
---|
| 12 | * HeuristicLab is distributed in the hope that it will be useful,
|
---|
| 13 | * but WITHOUT ANY WARRANTY; without even the implied warranty of
|
---|
| 14 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
---|
| 15 | * GNU General Public License for more details.
|
---|
| 16 | *
|
---|
| 17 | * You should have received a copy of the GNU General Public License
|
---|
| 18 | * along with HeuristicLab. If not, see <http://www.gnu.org/licenses/>.
|
---|
| 19 | */
|
---|
| 20 | #endregion
|
---|
| 21 |
|
---|
| 22 | using System;
|
---|
| 23 | using System.Collections.Generic;
|
---|
| 24 | using System.Linq;
|
---|
| 25 | using HeuristicLab.Common;
|
---|
| 26 | using HeuristicLab.Core;
|
---|
| 27 | using HeuristicLab.Persistence.Default.CompositeSerializers.Storable;
|
---|
| 28 | using HeuristicLab.Problems.DataAnalysis;
|
---|
| 29 | using MathNet.Numerics.Data.Matlab;
|
---|
| 30 | using MathNet.Numerics.LinearAlgebra;
|
---|
| 31 | using MathNet.Numerics.LinearAlgebra.Double;
|
---|
| 32 | using MathNet.Numerics.LinearAlgebra.Factorization;
|
---|
| 33 |
|
---|
| 34 | namespace HeuristicLab.Algorithms.DataAnalysis.Experimental {
|
---|
| 35 | // TODO: scale y
|
---|
| 36 | // TODO: remove dependence of scaling and export scaling parameters
|
---|
| 37 | // TODO: export / import all relevant data
|
---|
| 38 | [StorableClass]
|
---|
| 39 | [Item("GaussianProcessModelMKL", "Represents a Gaussian process posterior.")]
|
---|
| 40 | public sealed class GaussianProcessModelMKL : RegressionModel, IGaussianProcessModel {
|
---|
| 41 | public override IEnumerable<string> VariablesUsedForPrediction {
|
---|
| 42 | get { return allowedInputVariables; }
|
---|
| 43 | }
|
---|
| 44 |
|
---|
| 45 | [Storable]
|
---|
| 46 | private double negativeLogLikelihood;
|
---|
| 47 | public double NegativeLogLikelihood {
|
---|
| 48 | get { return negativeLogLikelihood; }
|
---|
| 49 | }
|
---|
| 50 |
|
---|
| 51 | [Storable]
|
---|
| 52 | private double negativeLooPredictiveProbability;
|
---|
| 53 | public double NegativeLooPredictiveProbability {
|
---|
| 54 | get { return negativeLooPredictiveProbability; }
|
---|
| 55 | }
|
---|
| 56 |
|
---|
| 57 | [Storable]
|
---|
| 58 | private double[] hyperparameterGradients;
|
---|
| 59 | public double[] HyperparameterGradients {
|
---|
| 60 | get {
|
---|
| 61 | var copy = new double[hyperparameterGradients.Length];
|
---|
| 62 | Array.Copy(hyperparameterGradients, copy, copy.Length);
|
---|
| 63 | return copy;
|
---|
| 64 | }
|
---|
| 65 | }
|
---|
| 66 |
|
---|
| 67 | [Storable]
|
---|
| 68 | private ICovarianceFunction covarianceFunction;
|
---|
| 69 | public ICovarianceFunction CovarianceFunction {
|
---|
| 70 | get { return covarianceFunction; }
|
---|
| 71 | }
|
---|
| 72 | [Storable]
|
---|
| 73 | private IMeanFunction meanFunction;
|
---|
| 74 | public IMeanFunction MeanFunction {
|
---|
| 75 | get { return meanFunction; }
|
---|
| 76 | }
|
---|
| 77 |
|
---|
| 78 | [Storable]
|
---|
| 79 | private string[] allowedInputVariables;
|
---|
| 80 | public string[] AllowedInputVariables {
|
---|
| 81 | get { return allowedInputVariables; }
|
---|
| 82 | }
|
---|
| 83 |
|
---|
| 84 | [Storable]
|
---|
| 85 | private Vector<double> alpha;
|
---|
| 86 | [Storable]
|
---|
| 87 | private double sqrSigmaNoise;
|
---|
| 88 | public double SigmaNoise {
|
---|
| 89 | get { return Math.Sqrt(sqrSigmaNoise); }
|
---|
| 90 | }
|
---|
| 91 |
|
---|
| 92 | [Storable]
|
---|
| 93 | private double[] meanParameter;
|
---|
| 94 | [Storable]
|
---|
| 95 | private double[] covarianceParameter;
|
---|
| 96 |
|
---|
| 97 | private Matrix<double> l;
|
---|
| 98 | private double[,] x; // scaled training dataset, used to be storable in previous versions (is calculated lazily now)
|
---|
| 99 |
|
---|
| 100 |
|
---|
| 101 | [Storable]
|
---|
| 102 | private IDataset trainingDataset; // it is better to store the original training dataset completely because this is more efficient in persistence
|
---|
| 103 | [Storable]
|
---|
| 104 | private int[] trainingRows;
|
---|
| 105 |
|
---|
| 106 | [Storable]
|
---|
| 107 | private Scaling inputScaling;
|
---|
| 108 |
|
---|
| 109 |
|
---|
| 110 | [StorableConstructor]
|
---|
| 111 | private GaussianProcessModelMKL(bool deserializing) : base(deserializing) { }
|
---|
| 112 | private GaussianProcessModelMKL(GaussianProcessModelMKL original, Cloner cloner)
|
---|
| 113 | : base(original, cloner) {
|
---|
| 114 | this.meanFunction = cloner.Clone(original.meanFunction);
|
---|
| 115 | this.covarianceFunction = cloner.Clone(original.covarianceFunction);
|
---|
| 116 | if (original.inputScaling != null)
|
---|
| 117 | this.inputScaling = cloner.Clone(original.inputScaling);
|
---|
| 118 | this.trainingDataset = cloner.Clone(original.trainingDataset);
|
---|
| 119 | this.negativeLogLikelihood = original.negativeLogLikelihood;
|
---|
| 120 | this.negativeLooPredictiveProbability = original.negativeLooPredictiveProbability;
|
---|
| 121 | this.sqrSigmaNoise = original.sqrSigmaNoise;
|
---|
| 122 | if (original.meanParameter != null) {
|
---|
| 123 | this.meanParameter = (double[])original.meanParameter.Clone();
|
---|
| 124 | }
|
---|
| 125 | if (original.covarianceParameter != null) {
|
---|
| 126 | this.covarianceParameter = (double[])original.covarianceParameter.Clone();
|
---|
| 127 | }
|
---|
| 128 |
|
---|
| 129 | // shallow copies of arrays because they cannot be modified
|
---|
| 130 | this.trainingRows = original.trainingRows;
|
---|
| 131 | this.allowedInputVariables = original.allowedInputVariables;
|
---|
| 132 | this.alpha = original.alpha;
|
---|
| 133 | this.l = original.l;
|
---|
| 134 | this.x = original.x;
|
---|
| 135 | }
|
---|
| 136 | public GaussianProcessModelMKL(IDataset ds, string targetVariable, IEnumerable<string> allowedInputVariables, IEnumerable<int> rows,
|
---|
| 137 | IEnumerable<double> hyp, IMeanFunction meanFunction, ICovarianceFunction covarianceFunction,
|
---|
| 138 | bool scaleInputs = true)
|
---|
| 139 | : base(targetVariable) {
|
---|
| 140 | // MathNet.Numerics.Control.UseNativeMKL();
|
---|
| 141 | // MathNet.Numerics.Control.UseSingleThread();
|
---|
| 142 | // this.Description += Control.LinearAlgebraProvider.ToString();
|
---|
| 143 | this.name = ItemName;
|
---|
| 144 | this.description = ItemDescription;
|
---|
| 145 | this.meanFunction = (IMeanFunction)meanFunction.Clone();
|
---|
| 146 | this.covarianceFunction = (ICovarianceFunction)covarianceFunction.Clone();
|
---|
| 147 | this.allowedInputVariables = allowedInputVariables.ToArray();
|
---|
| 148 |
|
---|
| 149 |
|
---|
| 150 | int nVariables = this.allowedInputVariables.Length;
|
---|
| 151 | meanParameter = hyp
|
---|
| 152 | .Take(this.meanFunction.GetNumberOfParameters(nVariables))
|
---|
| 153 | .ToArray();
|
---|
| 154 |
|
---|
| 155 | covarianceParameter = hyp.Skip(this.meanFunction.GetNumberOfParameters(nVariables))
|
---|
| 156 | .Take(this.covarianceFunction.GetNumberOfParameters(nVariables))
|
---|
| 157 | .ToArray();
|
---|
| 158 | sqrSigmaNoise = Math.Exp(2.0 * hyp.Last());
|
---|
| 159 | CalculateModel(ds, rows, scaleInputs);
|
---|
| 160 | }
|
---|
| 161 |
|
---|
| 162 | private void CalculateModel(IDataset ds, IEnumerable<int> rows, bool scaleInputs = true) {
|
---|
| 163 | this.trainingDataset = (IDataset)ds.Clone();
|
---|
| 164 | this.trainingRows = rows.ToArray();
|
---|
| 165 | this.inputScaling = scaleInputs ? new Scaling(ds, allowedInputVariables, rows) : null;
|
---|
| 166 |
|
---|
| 167 | x = GetData(ds, this.allowedInputVariables, this.trainingRows, this.inputScaling);
|
---|
| 168 |
|
---|
| 169 | IEnumerable<double> y;
|
---|
| 170 | y = ds.GetDoubleValues(TargetVariable, rows);
|
---|
| 171 |
|
---|
| 172 | int n = x.GetLength(0);
|
---|
| 173 |
|
---|
| 174 | var columns = Enumerable.Range(0, x.GetLength(1)).ToArray();
|
---|
| 175 | // calculate cholesky decomposed (lower triangular) covariance matrix
|
---|
| 176 | var cov = covarianceFunction.GetParameterizedCovarianceFunction(covarianceParameter, columns);
|
---|
| 177 | var chol = CalculateL(x, cov, sqrSigmaNoise);
|
---|
| 178 | this.l = chol.Factor;
|
---|
| 179 |
|
---|
| 180 | // calculate mean
|
---|
| 181 | var mean = meanFunction.GetParameterizedMeanFunction(meanParameter, columns);
|
---|
| 182 | double[] m = Enumerable.Range(0, x.GetLength(0))
|
---|
| 183 | .Select(r => mean.Mean(x, r))
|
---|
| 184 | .ToArray();
|
---|
| 185 |
|
---|
| 186 |
|
---|
| 187 | // solve for alpha
|
---|
| 188 | Vector<double> ym = DenseVector.OfEnumerable(y.Zip(m, (a, b) => a - b));
|
---|
| 189 |
|
---|
| 190 |
|
---|
| 191 | alpha = chol.Solve(ym);
|
---|
| 192 | alpha = alpha * 1.0 / sqrSigmaNoise;
|
---|
| 193 |
|
---|
| 194 | // calculate sum of diagonal elements for likelihood
|
---|
| 195 | double diagSum = chol.DeterminantLn;
|
---|
| 196 | negativeLogLikelihood = 0.5 * ym.DotProduct(alpha) + 0.5 * diagSum + (n / 2.0) * Math.Log(2.0 * Math.PI * sqrSigmaNoise);
|
---|
| 197 |
|
---|
| 198 | // derivatives
|
---|
| 199 | int nAllowedVariables = x.GetLength(1);
|
---|
| 200 |
|
---|
| 201 | Matrix<double> lCopy;
|
---|
| 202 |
|
---|
| 203 | lCopy = chol.Solve(DenseMatrix.CreateIdentity(n));
|
---|
| 204 |
|
---|
| 205 | // LOOCV log predictive probability (GPML page 116 and 117)
|
---|
| 206 | var sumLoo = 0.0;
|
---|
| 207 | var ki = new DenseVector(n);
|
---|
| 208 | for (int i = 0; i < n; i++) {
|
---|
| 209 | for (int j = 0; j < n; j++) ki[j] = cov.Covariance(x, i, j);
|
---|
| 210 | var yi = ki.DotProduct(alpha);
|
---|
| 211 | var yi_loo = yi - alpha[i] / lCopy[i, i] / sqrSigmaNoise;
|
---|
| 212 | var s2_loo = sqrSigmaNoise / lCopy[i, i];
|
---|
| 213 | var err = ym[i] - yi_loo;
|
---|
| 214 | var nll_loo = Math.Log(s2_loo) + err * err / s2_loo;
|
---|
| 215 | sumLoo += nll_loo;
|
---|
| 216 | }
|
---|
| 217 | sumLoo += n * Math.Log(2 * Math.PI);
|
---|
| 218 | negativeLooPredictiveProbability = 0.5 * sumLoo;
|
---|
| 219 |
|
---|
| 220 | for (int i = 0; i < n; i++) {
|
---|
| 221 | for (int j = 0; j <= i; j++)
|
---|
| 222 | lCopy[i, j] = lCopy[i, j] / sqrSigmaNoise - alpha[i] * alpha[j];
|
---|
| 223 | }
|
---|
| 224 |
|
---|
| 225 | double noiseGradient = sqrSigmaNoise * Enumerable.Range(0, n).Select(i => lCopy[i, i]).Sum();
|
---|
| 226 |
|
---|
| 227 | double[] meanGradients = new double[meanFunction.GetNumberOfParameters(nAllowedVariables)];
|
---|
| 228 | for (int k = 0; k < meanGradients.Length; k++) {
|
---|
| 229 | var meanGrad = new DenseVector(alpha.Count);
|
---|
| 230 | for (int g = 0; g < meanGrad.Count; g++)
|
---|
| 231 | meanGrad[g] = mean.Gradient(x, g, k);
|
---|
| 232 |
|
---|
| 233 | meanGradients[k] = -meanGrad.DotProduct(alpha);
|
---|
| 234 | }
|
---|
| 235 |
|
---|
| 236 | double[] covGradients = new double[covarianceFunction.GetNumberOfParameters(nAllowedVariables)];
|
---|
| 237 | if (covGradients.Length > 0) {
|
---|
| 238 | for (int i = 0; i < n; i++) {
|
---|
| 239 | for (int j = 0; j < i; j++) {
|
---|
| 240 | var g = cov.CovarianceGradient(x, i, j);
|
---|
| 241 | for (int k = 0; k < covGradients.Length; k++) {
|
---|
| 242 | covGradients[k] += lCopy[i, j] * g[k];
|
---|
| 243 | }
|
---|
| 244 | }
|
---|
| 245 |
|
---|
| 246 | var gDiag = cov.CovarianceGradient(x, i, i);
|
---|
| 247 | for (int k = 0; k < covGradients.Length; k++) {
|
---|
| 248 | // diag
|
---|
| 249 | covGradients[k] += 0.5 * lCopy[i, i] * gDiag[k];
|
---|
| 250 | }
|
---|
| 251 | }
|
---|
| 252 | }
|
---|
| 253 |
|
---|
| 254 | hyperparameterGradients =
|
---|
| 255 | meanGradients
|
---|
| 256 | .Concat(covGradients)
|
---|
| 257 | .Concat(new double[] { noiseGradient }).ToArray();
|
---|
| 258 |
|
---|
| 259 | }
|
---|
| 260 |
|
---|
| 261 | private static double[,] GetData(IDataset ds, IEnumerable<string> allowedInputs, IEnumerable<int> rows, Scaling scaling) {
|
---|
| 262 | if (scaling != null) {
|
---|
| 263 | // BackwardsCompatibility3.3
|
---|
| 264 | #region Backwards compatible code, remove with 3.4
|
---|
| 265 | // TODO: completely remove Scaling class
|
---|
| 266 | List<string> variablesList = allowedInputs.ToList();
|
---|
| 267 | List<int> rowsList = rows.ToList();
|
---|
| 268 |
|
---|
| 269 | double[,] matrix = new double[rowsList.Count, variablesList.Count];
|
---|
| 270 |
|
---|
| 271 | int col = 0;
|
---|
| 272 | foreach (string column in variablesList) {
|
---|
| 273 | var values = scaling.GetScaledValues(ds, column, rowsList);
|
---|
| 274 | int row = 0;
|
---|
| 275 | foreach (var value in values) {
|
---|
| 276 | matrix[row, col] = value;
|
---|
| 277 | row++;
|
---|
| 278 | }
|
---|
| 279 | col++;
|
---|
| 280 | }
|
---|
| 281 | return matrix;
|
---|
| 282 | #endregion
|
---|
| 283 | } else {
|
---|
| 284 | return ds.ToArray(allowedInputs, rows);
|
---|
| 285 | }
|
---|
| 286 | }
|
---|
| 287 |
|
---|
| 288 | private static Cholesky<double> CalculateL(double[,] x, ParameterizedCovarianceFunction cov, double sqrSigmaNoise) {
|
---|
| 289 | int n = x.GetLength(0);
|
---|
| 290 | var l = new DenseMatrix(n, n);
|
---|
| 291 |
|
---|
| 292 | // calculate covariances
|
---|
| 293 | for (int i = 0; i < n; i++) {
|
---|
| 294 | for (int j = i; j < n; j++) {
|
---|
| 295 | l[j, i] = l[i, j] = cov.Covariance(x, i, j) / sqrSigmaNoise;
|
---|
| 296 | if (j == i) l[j, i] += 1.0;
|
---|
| 297 | }
|
---|
| 298 | }
|
---|
| 299 |
|
---|
| 300 | // cholesky decomposition
|
---|
| 301 | return l.Cholesky();
|
---|
| 302 | }
|
---|
| 303 |
|
---|
| 304 |
|
---|
| 305 | public override IDeepCloneable Clone(Cloner cloner) {
|
---|
| 306 | return new GaussianProcessModelMKL(this, cloner);
|
---|
| 307 | }
|
---|
| 308 |
|
---|
| 309 | // is called by the solution creator to set all parameter values of the covariance and mean function
|
---|
| 310 | // to the optimized values (necessary to make the values visible in the GUI)
|
---|
| 311 | public void FixParameters() {
|
---|
| 312 | covarianceFunction.SetParameter(covarianceParameter);
|
---|
| 313 | meanFunction.SetParameter(meanParameter);
|
---|
| 314 | covarianceParameter = new double[0];
|
---|
| 315 | meanParameter = new double[0];
|
---|
| 316 | }
|
---|
| 317 |
|
---|
| 318 | #region IRegressionModel Members
|
---|
| 319 | public override IEnumerable<double> GetEstimatedValues(IDataset dataset, IEnumerable<int> rows) {
|
---|
| 320 | return GetEstimatedValuesHelper(dataset, rows);
|
---|
| 321 | }
|
---|
| 322 | public override IRegressionSolution CreateRegressionSolution(IRegressionProblemData problemData) {
|
---|
| 323 | return new GaussianProcessRegressionSolution(this, new RegressionProblemData(problemData));
|
---|
| 324 | }
|
---|
| 325 | #endregion
|
---|
| 326 |
|
---|
| 327 |
|
---|
| 328 | private IEnumerable<double> GetEstimatedValuesHelper(IDataset dataset, IEnumerable<int> rows) {
|
---|
| 329 | if (x == null) {
|
---|
| 330 | x = GetData(trainingDataset, allowedInputVariables, trainingRows, inputScaling);
|
---|
| 331 | }
|
---|
| 332 | int n = x.GetLength(0);
|
---|
| 333 |
|
---|
| 334 | double[,] newX = GetData(dataset, allowedInputVariables, rows, inputScaling);
|
---|
| 335 | int newN = newX.GetLength(0);
|
---|
| 336 |
|
---|
| 337 | var columns = Enumerable.Range(0, newX.GetLength(1)).ToArray();
|
---|
| 338 | var mean = meanFunction.GetParameterizedMeanFunction(meanParameter, columns);
|
---|
| 339 | var ms = Enumerable.Range(0, newX.GetLength(0))
|
---|
| 340 | .Select(r => mean.Mean(newX, r))
|
---|
| 341 | .ToArray();
|
---|
| 342 | var cov = covarianceFunction.GetParameterizedCovarianceFunction(covarianceParameter, columns);
|
---|
| 343 | for (int i = 0; i < newN; i++) {
|
---|
| 344 | var Ks = new DenseVector(n);
|
---|
| 345 | for (int j = 0; j < n; j++) {
|
---|
| 346 | Ks[j] = cov.CrossCovariance(x, newX, j, i);
|
---|
| 347 | }
|
---|
| 348 | yield return ms[i] + Ks.DotProduct(alpha);
|
---|
| 349 | }
|
---|
| 350 | }
|
---|
| 351 |
|
---|
| 352 | public IEnumerable<double> GetEstimatedVariances(IDataset dataset, IEnumerable<int> rows) {
|
---|
| 353 | if (x == null) {
|
---|
| 354 | x = GetData(trainingDataset, allowedInputVariables, trainingRows, inputScaling);
|
---|
| 355 | }
|
---|
| 356 | int n = x.GetLength(0);
|
---|
| 357 |
|
---|
| 358 | var newX = GetData(dataset, allowedInputVariables, rows, inputScaling);
|
---|
| 359 | int newN = newX.GetLength(0);
|
---|
| 360 | if (newN == 0) return Enumerable.Empty<double>();
|
---|
| 361 |
|
---|
| 362 | var kss = new DenseVector(newN);
|
---|
| 363 | Matrix<double> sWKs = new DenseMatrix(n, newN);
|
---|
| 364 | var columns = Enumerable.Range(0, newX.GetLength(1)).ToArray();
|
---|
| 365 | var cov = covarianceFunction.GetParameterizedCovarianceFunction(covarianceParameter, columns);
|
---|
| 366 |
|
---|
| 367 | if (l == null) {
|
---|
| 368 | l = CalculateL(x, cov, sqrSigmaNoise).Factor;
|
---|
| 369 | }
|
---|
| 370 |
|
---|
| 371 | // for stddev
|
---|
| 372 | for (int i = 0; i < newN; i++)
|
---|
| 373 | kss[i] = cov.Covariance(newX, i, i);
|
---|
| 374 |
|
---|
| 375 | for (int i = 0; i < newN; i++) {
|
---|
| 376 | for (int j = 0; j < n; j++) {
|
---|
| 377 | sWKs[j, i] = cov.CrossCovariance(x, newX, j, i) / Math.Sqrt(sqrSigmaNoise);
|
---|
| 378 | }
|
---|
| 379 | }
|
---|
| 380 |
|
---|
| 381 | sWKs = l.Solve(sWKs);
|
---|
| 382 |
|
---|
| 383 | for (int i = 0; i < newN; i++) {
|
---|
| 384 | var col = sWKs.Column(i);
|
---|
| 385 | var sumV = col.DotProduct(col);
|
---|
| 386 | kss[i] += sqrSigmaNoise; // kss is V(f), add noise variance of predictive distibution to get V(y)
|
---|
| 387 | kss[i] -= sumV;
|
---|
| 388 | if (kss[i] < 0) kss[i] = 0;
|
---|
| 389 | }
|
---|
| 390 | return kss;
|
---|
| 391 | }
|
---|
| 392 |
|
---|
| 393 |
|
---|
| 394 | public void Export(string fileName) {
|
---|
| 395 | MatlabWriter.Write<double>(fileName,
|
---|
| 396 | new Matrix<double>[] { DenseMatrix.OfArray(x), l, alpha.ToRowMatrix()},
|
---|
| 397 | new string[] { "x", "l", "alpha", }
|
---|
| 398 | );
|
---|
| 399 | }
|
---|
| 400 | }
|
---|
| 401 | }
|
---|