Changeset 13750 for branches/PerformanceComparison/HeuristicLab.OptimizationExpertSystem/3.3/Views

Timestamp:

04/08/16 14:24:51 (9 years ago)

Author:

abeham

Message:

#2457: Added SOM projection for problem instances, fixed a bug (learningRadius was not used)

File:

• branches/PerformanceComparison/HeuristicLab.OptimizationExpertSystem/3.3/Views/UnderstandingSolutionsView.cs (modified) (3 diffs)

branches/PerformanceComparison/HeuristicLab.OptimizationExpertSystem/3.3/Views/UnderstandingSolutionsView.cs

@@ -22 +22 @@
 using HeuristicLab.Analysis;
 using HeuristicLab.Analysis.QualityAnalysis;
+using HeuristicLab.Analysis.SelfOrganizingMaps;
 using HeuristicLab.Common;
 using HeuristicLab.Core;
@@ -241 +242 @@
         DoubleMatrix coords = null;
         if (projection == "SOM")
-          coords = Som(dissimilarities, new MersenneTwister(42), jittering: true);
+          coords = RelationalSOM.Map(dissimilarities, new MersenneTwister(42), jittering: true);
         else coords = MultidimensionalScaling.KruskalShepard(dissimilarities);
         var dataPoints = new List<DataPoint>();
@@ -382 +383 @@
       return dval.Value;
     }
-
-    #region SOM projection
-    /// <summary>
-    /// This is the online algorithm described in
-    /// Olteanu, M. and Villa-Vialaneix, N. 2015.
-    /// On-line relational and multiple relational SOM.
-    /// Neurocomputing 147, pp. 15-30.
-    /// </summary>
-    /// <param name="dissimilarities">The full NxN matrix containing all dissimilarities between N points.</param>
-    /// <param name="random">The random number generator to use.</param>
-    /// <param name="somSize">The length of a side of the SOM grid (there are somSize * somSize neurons).</param>
-    /// <param name="iterations">The amount of iterations to perform in learning.</param>
-    /// <param name="learningRate">The initial learning rate.</param>
-    /// <param name="learningRadius">The initial learning radius.</param>
-    /// <param name="jittering">If the final coordinates should be jittered slightly within the grid.</param>
-    /// <returns>A matrix of coordinates having N rows and 2 columns.</returns>
-    private DoubleMatrix Som(DoubleMatrix dissimilarities, IRandom random, int somSize = 5, int iterations = 100, double learningRate = double.NaN, double learningRadius = 5.0, bool jittering = true) {
-      var K = somSize * somSize;
-      var N = dissimilarities.Rows;
-      if (double.IsNaN(learningRate)) learningRate = 1.0 / Math.Sqrt(2.0 * N);
-      var fixedLearningRate = learningRate / 10.0;
-      var varLearningRate = 9.0 * fixedLearningRate;
-      Func<int, int, double> learningRateT = (maxIter, iter) => {
-        return varLearningRate * ((maxIter - iter) / (double)maxIter) + fixedLearningRate;
-      };
-      Func<int, int> getX = (neuron) => neuron % somSize;
-      Func<int, int> getY = (neuron) => neuron / somSize;
-      Func<int, int, int, int, double> neighborhood = (maxIter, iter, k, bmu) => {
-        var sigma = 1.0 * ((maxIter - iter) / (double)maxIter) + 0.0001;
-        var xK = getX(k);
-        var yK = getY(k);
-        var xW = getX(bmu);
-        var yW = getY(bmu);
-        var d = (xK - xW) * (xK - xW) + (yK - yW) * (yK - yW);
-        return Math.Exp(-d / (2.0 * sigma * sigma));
-      };
-      var alphas = Enumerable.Range(0, K).Select(k => Enumerable.Range(0, N).Select(_ => random.NextDouble()).ToArray()).ToArray();
-      // normalize s.t. sum(alphas[k]) = 1
-      for (var k = 0; k < K; k++) {
-        var sum = alphas[k].Sum();
-        for (var i = 0; i < alphas[k].Length; i++) alphas[k][i] /= sum;
-      }
-      var oldAlphas = alphas.Select(x => (double[])x.Clone()).ToArray();
-
-      for (var iter = 0; iter < iterations; iter++) {
-        var pointShuffle = Enumerable.Range(0, N).Shuffle(random).ToArray();
-        for (var p = 0; p < N; p++) {
-          var i = pointShuffle[p];
-          var bmu = GetBestMatchingUnit(dissimilarities, alphas, i);
-
-          for (var k = 0; k < K; k++) {
-            for (var j = 0; j < N; j++) {
-              alphas[k][j] = oldAlphas[k][j] + learningRateT(iterations, iter) * neighborhood(iterations, iter, k, bmu) * ((i == j ? 1.0 : 0.0) - oldAlphas[k][j]);
-            }
-          }
-        }
-        for (var k = 0; k < K; k++) {
-          for (var j = 0; j < N; j++) {
-            oldAlphas[k][j] = alphas[k][j];
-          }
-        }
-      }
-
-      var result = new DoubleMatrix(N, 2);
-      for (var i = 0; i < N; i++) {
-        var bmu = GetBestMatchingUnit(dissimilarities, alphas, i);
-        if (!jittering) {
-          result[i, 0] = getX(bmu);
-          result[i, 1] = getY(bmu);
-        } else {
-          result[i, 0] = getX(bmu) + random.NextDouble() * 0.8;
-          result[i, 1] = getY(bmu) + random.NextDouble() * 0.8;
-        }
-      }
-      return result;
-    }
-
-    private int GetBestMatchingUnit(DoubleMatrix D, double[][] alphas, int i) {
-      var bmu = -1;
-      var minV = double.MaxValue;
-      for (var k = 0; k < alphas.Length; k++) {
-        var Daki = 0.0;
-        var akDak = 0.0;
-        for (var r = 0; r < D.Rows; r++) {
-          var Dakr = 0.0;
-          for (var s = 0; s < D.Rows; s++) {
-            Dakr += D[r, s] * alphas[k][s];
-          }
-          if (r == i) Daki = Dakr;
-          akDak += alphas[k][r] * Dakr;
-        }
-        var v = Daki - 0.5 * akDak;
-        if (v < minV) {
-          bmu = k;
-          minV = v;
-        }
-      }
-      return bmu;
-    }
-    #endregion
-
-    private DoubleMatrix Mds(DoubleMatrix dissimilarities) {
-      return MultidimensionalScaling.KruskalShepard(dissimilarities);
-    }
     #endregion
   }