1 | #region License Information
|
---|
2 | /* HeuristicLab
|
---|
3 | * Copyright (C) 2002-2018 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.Core;
|
---|
26 | using HeuristicLab.Problems.DataAnalysis;
|
---|
27 | using HeuristicLab.Random;
|
---|
28 |
|
---|
29 | namespace HeuristicLab.Problems.Instances.DataAnalysis {
|
---|
30 | public sealed class GaussianProcessVariableNetwork : VariableNetwork {
|
---|
31 | private int numberOfFeatures;
|
---|
32 | private double noiseRatio;
|
---|
33 |
|
---|
34 | public override string Name { get { return string.Format("GaussianProcessVariableNetwork-{0:0%} ({1} dim)", noiseRatio, numberOfFeatures); } }
|
---|
35 |
|
---|
36 | public GaussianProcessVariableNetwork(int numberOfFeatures, double noiseRatio,
|
---|
37 | IRandom rand)
|
---|
38 | : base(250, 250, numberOfFeatures, noiseRatio, rand) {
|
---|
39 | this.noiseRatio = noiseRatio;
|
---|
40 | this.numberOfFeatures = numberOfFeatures;
|
---|
41 | }
|
---|
42 |
|
---|
43 | // sample the input variables that are actually used and sample from a Gaussian process
|
---|
44 | protected override IEnumerable<double> GenerateRandomFunction(IRandom rand, List<List<double>> xs, out string[] selectedVarNames, out double[] relevance) {
|
---|
45 | int nl = SampleNumberOfVariables(rand, xs.Count);
|
---|
46 |
|
---|
47 | var selectedIdx = Enumerable.Range(0, xs.Count).Shuffle(rand)
|
---|
48 | .Take(nl).ToArray();
|
---|
49 |
|
---|
50 | var selectedVars = selectedIdx.Select(i => xs[i]).ToArray();
|
---|
51 | selectedVarNames = selectedIdx.Select(i => VariableNames[i]).ToArray();
|
---|
52 | return SampleGaussianProcess(rand, selectedVars, out relevance);
|
---|
53 | }
|
---|
54 |
|
---|
55 | private IEnumerable<double> SampleGaussianProcess(IRandom rand, List<double>[] xs, out double[] relevance) {
|
---|
56 | int nl = xs.Length;
|
---|
57 | int nRows = xs.First().Count;
|
---|
58 |
|
---|
59 | // sample u iid ~ N(0, 1)
|
---|
60 | var u = Enumerable.Range(0, nRows).Select(_ => NormalDistributedRandom.NextDouble(rand, 0, 1)).ToArray();
|
---|
61 |
|
---|
62 | // sample actual length-scales
|
---|
63 | var l = Enumerable.Range(0, nl)
|
---|
64 | .Select(_ => rand.NextDouble() * 2 + 0.5)
|
---|
65 | .ToArray();
|
---|
66 |
|
---|
67 | double[,] K = CalculateCovariance(xs, l);
|
---|
68 |
|
---|
69 | // decompose
|
---|
70 | alglib.trfac.spdmatrixcholesky(ref K, nRows, false);
|
---|
71 |
|
---|
72 |
|
---|
73 | // calc y = Lu
|
---|
74 | var y = new double[u.Length];
|
---|
75 | alglib.ablas.rmatrixmv(nRows, nRows, K, 0, 0, 0, u, 0, ref y, 0);
|
---|
76 |
|
---|
77 | // calculate relevance by removing dimensions
|
---|
78 | relevance = CalculateRelevance(y, u, xs, l);
|
---|
79 |
|
---|
80 | return y;
|
---|
81 | }
|
---|
82 |
|
---|
83 | // calculate variable relevance based on removal of variables
|
---|
84 | // 1) to remove a variable we set it's length scale to infinity (no relation of the variable value to the target)
|
---|
85 | // 2) calculate MSE of the original target values (y) to the updated targes y' (after variable removal)
|
---|
86 | // 3) relevance is larger if MSE(y,y') is large
|
---|
87 | // 4) scale impacts so that the most important variable has impact = 1
|
---|
88 | private double[] CalculateRelevance(double[] y, double[] u, List<double>[] xs, double[] l) {
|
---|
89 | int nRows = xs.First().Count;
|
---|
90 | var changedL = new double[l.Length];
|
---|
91 | var relevance = new double[l.Length];
|
---|
92 | for(int i = 0; i < l.Length; i++) {
|
---|
93 | Array.Copy(l, changedL, changedL.Length);
|
---|
94 | changedL[i] = double.MaxValue;
|
---|
95 | var changedK = CalculateCovariance(xs, changedL);
|
---|
96 |
|
---|
97 | var yChanged = new double[u.Length];
|
---|
98 | alglib.ablas.rmatrixmv(nRows, nRows, changedK, 0, 0, 0, u, 0, ref yChanged, 0);
|
---|
99 |
|
---|
100 | OnlineCalculatorError error;
|
---|
101 | var mse = OnlineMeanSquaredErrorCalculator.Calculate(y, yChanged, out error);
|
---|
102 | if(error != OnlineCalculatorError.None) mse = double.MaxValue;
|
---|
103 | relevance[i] = mse;
|
---|
104 | }
|
---|
105 | // scale so that max relevance is 1.0
|
---|
106 | var maxRel = relevance.Max();
|
---|
107 | for(int i = 0; i < relevance.Length; i++) relevance[i] /= maxRel;
|
---|
108 | return relevance;
|
---|
109 | }
|
---|
110 |
|
---|
111 | private double[,] CalculateCovariance(List<double>[] xs, double[] l) {
|
---|
112 | int nRows = xs.First().Count;
|
---|
113 | double[,] K = new double[nRows, nRows];
|
---|
114 | for(int r = 0; r < nRows; r++) {
|
---|
115 | double[] xi = xs.Select(x => x[r]).ToArray();
|
---|
116 | for(int c = 0; c <= r; c++) {
|
---|
117 | double[] xj = xs.Select(x => x[c]).ToArray();
|
---|
118 | double dSqr = xi.Zip(xj, (xik, xjk) => (xik - xjk))
|
---|
119 | .Select(dk => dk * dk)
|
---|
120 | .Zip(l, (dk, lk) => dk / lk)
|
---|
121 | .Sum();
|
---|
122 | K[r, c] = Math.Exp(-dSqr);
|
---|
123 | }
|
---|
124 | }
|
---|
125 | // add a small diagonal matrix for numeric stability
|
---|
126 | for(int i = 0; i < nRows; i++) {
|
---|
127 | K[i, i] += 1.0E-7;
|
---|
128 | }
|
---|
129 |
|
---|
130 | return K;
|
---|
131 | }
|
---|
132 | }
|
---|
133 | }
|
---|