- Timestamp:
- 11/29/16 15:46:48 (8 years ago)
- Location:
- branches/PerformanceComparison/HeuristicLab.OptimizationExpertSystem.Common/3.3
- Files:
-
- 1 edited
- 1 moved
Legend:
- Unmodified
- Added
- Removed
-
branches/PerformanceComparison/HeuristicLab.OptimizationExpertSystem.Common/3.3/HeuristicLab.OptimizationExpertSystem.Common-3.3.csproj
r13861 r14429 164 164 <ItemGroup> 165 165 <Compile Include="Interfaces\IRecommendationModel.cs" /> 166 <Compile Include="ProblemCharacteristicAnalysis\QAP\QAP BestDirectedWalk.cs" />166 <Compile Include="ProblemCharacteristicAnalysis\QAP\QAPDirectedWalk.cs" /> 167 167 <Compile Include="Recommenders\KNearestNeighborModel.cs" /> 168 168 <Compile Include="Recommenders\OverallBestRecommender.cs" /> -
branches/PerformanceComparison/HeuristicLab.OptimizationExpertSystem.Common/3.3/ProblemCharacteristicAnalysis/QAP/QAPDirectedWalk.cs
r14428 r14429 85 85 86 86 public override IEnumerable<IResult> Calculate() { 87 IRandom random = Seed.HasValue ? new MersenneTwister((uint)Seed.Value) : new MersenneTwister(); 88 var qap = (QuadraticAssignmentProblem)Problem; 87 89 var pathCount = Paths; 88 var random = Seed.HasValue ? new MersenneTwister((uint)Seed.Value) : new MersenneTwister(); 89 var bestImprovement = BestImprovement; 90 var qap = (QuadraticAssignmentProblem)Problem; 91 var N = qap.Weights.Rows; 92 var start = new Permutation(PermutationTypes.Absolute, N, random); 93 var end = start; 94 List<double> ips = new List<double>(), ups = new List<double>(), sd2 = new List<double>(); 95 var pathsD1 = new List<List<Tuple<double, double>>>(); 96 for (var i = 0; i < pathCount; i++) { 97 var trajD1 = new List<Tuple<double, double>>(); 98 pathsD1.Add(trajD1); 99 100 if ((i + 1) % N == 0) { 101 start = new Permutation(PermutationTypes.Absolute, N, random); 102 end = start; 103 } 104 end = (Permutation)end.Clone(); 105 Rot1(end); 106 107 var hist = new Tuple<Permutation, double>[5]; 108 var walkDist = 0.0; 109 var prevVal = double.NaN; 110 var sumD2 = 0.0; 111 var inflectionPoints = 0; 112 var undulationPoints = 0; 113 var countPoints = 0; 114 var counter = 0; 115 var path = bestImprovement ? BestImprovementWalk(qap, start, QAPEvaluator.Apply(start, qap.Weights, qap.Distances), end) 116 : FirstImprovementWalk(qap, start, QAPEvaluator.Apply(start, qap.Weights, qap.Distances), end, random); 117 foreach (var next in path) { 118 if (hist[0] != null) { 119 var dist = Dist(next.Item1, hist[0].Item1); 120 walkDist += dist; 90 91 var perm = new Permutation(PermutationTypes.Absolute, qap.Weights.Rows, random); 92 var permutations = new List<Permutation> { perm }; 93 while (permutations.Count < pathCount + 1) { 94 perm = (Permutation)permutations.Last().Clone(); 95 BiasedShuffle(perm, random); 96 permutations.Add(perm); 97 } 98 99 var trajectories = Run(random, (QuadraticAssignmentProblem)Problem, permutations, BestImprovement).ToList(); 100 var firstDerivatives = trajectories.Select(path => ApproximateDerivative(path).ToList()).ToList(); 101 var secondDerivatives = firstDerivatives.Select(d1 => ApproximateDerivative(d1).ToList()).ToList(); 102 103 var props = GetCharacteristics(trajectories, firstDerivatives, secondDerivatives).ToDictionary(x => x.Item1, x => x.Item2); 104 foreach (var chara in characteristics.CheckedItems.Select(x => x.Value.Value)) { 105 if (chara == "Swap2.Sharpness") yield return new Result("Swap2.Sharpness", new DoubleValue(props["Sharpness"])); 106 if (chara == "Swap2.Bumpiness") yield return new Result("Swap2.Bumpiness", new DoubleValue(props["Bumpiness"])); 107 if (chara == "Swap2.Flatness") yield return new Result("Swap2.Flatness", new DoubleValue(props["Flatness"])); 108 if (chara == "Swap2.Steadiness") yield return new Result("Swap2.Steadiness", new DoubleValue(props["Steadiness"])); 109 } 110 } 111 112 public static IEnumerable<List<Tuple<Permutation, double>>> Run(IRandom random, QuadraticAssignmentProblem qap, IEnumerable<Permutation> permutations, bool bestImprovement = true) { 113 var iter = permutations.GetEnumerator(); 114 if (!iter.MoveNext()) yield break; 115 116 var start = iter.Current; 117 while (iter.MoveNext()) { 118 var end = iter.Current; 119 120 var walk = (bestImprovement ? BestDirectedWalk(qap, start, end) : FirstDirectedWalk(random, qap, start, end)).ToList(); 121 var max = walk.Max(x => x.Item2); 122 var min = walk.Min(x => x.Item2); 123 if (max > min) 124 yield return walk.Select(x => Tuple.Create(x.Item1, (x.Item2 - min) / (max - min))).ToList(); 125 else yield return walk.Select(x => Tuple.Create(x.Item1, 0.0)).ToList(); 126 start = end; 127 } // end paths 128 } 129 130 private IEnumerable<Tuple<string, double>> GetCharacteristics(List<List<Tuple<Permutation, double>>> f, List<List<Tuple<Permutation, double>>> f1, List<List<Tuple<Permutation, double>>> f2) { 131 var sharpness = f2.Average(x => Area(x)); 132 var bumpiness = 0.0; 133 var flatness = 0.0; 134 var downPointing = f1.Where(x => x.Min(y => y.Item2) < 0).ToList(); 135 136 var steadiness = 0.0; 137 foreach (var path in downPointing) { 138 steadiness += ComBelowZero(path); 139 } 140 if (downPointing.Count > 0) steadiness /= downPointing.Count; 141 142 for (var p = 0; p < f2.Count; p++) { 143 var bump = 0; 144 var flat = 0; 145 for (var i = 0; i < f2[p].Count - 1; i++) { 146 if ((f2[p][i].Item2 > 0 && f2[p][i + 1].Item2 < 0) || (f2[p][i].Item2 < 0 && f2[p][i + 1].Item2 > 0)) { 147 bump++; 148 } else if (f2[p][i].Item2 == 0) { 149 flat++; 121 150 } 122 // from the past 5 values we can calculate the 2nd derivative 123 // first derivative in point 2 as differential between points 1 and 3 124 // first derivative in point 4 as differential between points 3 and 5 125 // second derivative in point 3 as differential between the first derivatives in points 2 and 4 126 hist[4] = hist[3]; 127 hist[3] = hist[2]; 128 hist[2] = hist[1]; 129 hist[1] = hist[0]; 130 hist[0] = next; 131 counter++; 132 133 if (counter < 3) continue; 134 var grad1 = (hist[0].Item2 - hist[2].Item2) / Dist(hist[0].Item1, hist[2].Item1); 135 136 if (!double.IsNaN(grad1) && !double.IsInfinity(grad1)) 137 trajD1.Add(Tuple.Create(walkDist, grad1)); 138 139 if (counter < 5) continue; 140 countPoints++; 141 var grad2 = (hist[2].Item2 - hist[4].Item2) / Dist(hist[2].Item1, hist[4].Item1); 142 var dgrad = (grad1 - grad2) / Dist(hist[1].Item1, hist[3].Item1); 143 144 if (double.IsNaN(dgrad) || double.IsInfinity(dgrad)) continue; 145 if (!double.IsNaN(prevVal)) { 146 if (prevVal < 0 && dgrad > 0 || prevVal > 0 && dgrad < 0) 147 inflectionPoints++; 148 else if (prevVal.IsAlmost(0) && dgrad.IsAlmost(0) 149 || prevVal.IsAlmost(0) && !dgrad.IsAlmost(0) 150 || !prevVal.IsAlmost(0) && dgrad.IsAlmost(0)) 151 undulationPoints++; 152 } 153 sumD2 += Math.Abs(grad1 - grad2); 154 prevVal = dgrad; 155 } 156 start = end; 157 ips.Add(inflectionPoints / (double)countPoints); 158 ups.Add(undulationPoints / (double)countPoints); 159 sd2.Add(sumD2 / walkDist); 160 } // end paths 161 var avgZero = pathsD1.Select(path => path.SkipWhile(v => v.Item2 < 0).First().Item1 / path.Last().Item1).Median(); 162 163 foreach (var chara in characteristics.CheckedItems.Select(x => x.Value.Value)) { 164 if (chara == "Swap2.Sharpness") yield return new Result("Swap2.Sharpness", new DoubleValue(sd2.Average())); 165 if (chara == "Swap2.Bumpiness") yield return new Result("Swap2.Bumpiness", new DoubleValue(ips.Average())); 166 if (chara == "Swap2.Flatness") yield return new Result("Swap2.Flatness", new DoubleValue(ups.Average())); 167 if (chara == "Swap2.Steadiness") yield return new Result("Swap2.Steadiness", new DoubleValue(avgZero)); 168 } 169 } 170 171 public IEnumerable<Tuple<Permutation, double>> BestImprovementWalk(QuadraticAssignmentProblem qap, Permutation start, double fitness, Permutation end) { 151 } 152 bumpiness += bump / (f2[p].Count - 1.0); 153 flatness += flat / (f2[p].Count - 1.0); 154 } 155 bumpiness /= f2.Count; 156 flatness /= f2.Count; 157 return new[] { 158 Tuple.Create("Sharpness", sharpness), 159 Tuple.Create("Bumpiness", bumpiness), 160 Tuple.Create("Flatness", flatness), 161 Tuple.Create("Steadiness", steadiness) 162 }; 163 } 164 165 public static IEnumerable<Tuple<Permutation, double>> BestDirectedWalk(QuadraticAssignmentProblem qap, Permutation start, Permutation end) { 172 166 var N = qap.Weights.Rows; 173 167 var sol = start; 168 var fitness = QAPEvaluator.Apply(sol, qap.Weights, qap.Distances); 169 yield return Tuple.Create(sol, fitness); 170 174 171 var invSol = GetInverse(sol); 175 172 // we require at most N-1 steps to move from one permutation to another … … 197 194 } 198 195 199 public IEnumerable<Tuple<Permutation, double>> FirstImprovementWalk(QuadraticAssignmentProblem qap, Permutation start, double fitness, Permutation end, IRandom random) {196 public static IEnumerable<Tuple<Permutation, double>> FirstDirectedWalk(IRandom random, QuadraticAssignmentProblem qap, Permutation start, Permutation end) { 200 197 var N = qap.Weights.Rows; 201 198 var sol = start; 199 var fitness = QAPEvaluator.Apply(sol, qap.Weights, qap.Distances); 200 yield return Tuple.Create(sol, fitness); 201 202 202 var invSol = GetInverse(sol); 203 203 // randomize the order in which improvements are tried … … 229 229 } 230 230 231 private static double Area(IEnumerable<Tuple<Permutation, double>> path) { 232 var iter = path.GetEnumerator(); 233 if (!iter.MoveNext()) return 0.0; 234 var area = 0.0; 235 var prev = iter.Current; 236 while (iter.MoveNext()) { 237 area += TrapezoidArea(prev, iter.Current); 238 prev = iter.Current; 239 } 240 return area; 241 } 242 243 private static double TrapezoidArea(Tuple<Permutation, double> a, Tuple<Permutation, double> b) { 244 var area = 0.0; 245 var dist = Dist(a.Item1, b.Item1); 246 if ((a.Item2 <= 0 && b.Item2 <= 0) || (a.Item2 >= 0 && b.Item2 >= 0)) 247 area += dist * (Math.Abs(a.Item2) + Math.Abs(b.Item2)) / 2.0; 248 else { 249 var k = (b.Item2 - a.Item2) / dist; 250 var d = a.Item2; 251 var x = -d / k; 252 area += Math.Abs(x * a.Item2 / 2.0); 253 area += Math.Abs((dist - x) * b.Item2 / 2.0); 254 } 255 return area; 256 } 257 258 private static double ComBelowZero(IEnumerable<Tuple<Permutation, double>> path) { 259 var area = 0.0; 260 var com = 0.0; 261 var nwalkDist = 0.0; 262 Tuple<Permutation, double> prev = null; 263 var iter = path.GetEnumerator(); 264 while (iter.MoveNext()) { 265 var c = iter.Current; 266 if (prev != null) { 267 var ndist = Dist(prev.Item1, c.Item1) / (double)c.Item1.Length; 268 nwalkDist += ndist; 269 if (prev.Item2 < 0 || c.Item2 < 0) { 270 var a = TrapezoidArea(prev, c) / (double)c.Item1.Length; 271 area += a; 272 com += (nwalkDist - (ndist / 2.0)) * a; 273 } 274 } 275 prev = c; 276 } 277 return com / area; 278 } 279 280 private static IEnumerable<Tuple<Permutation, double>> ApproximateDerivative(IEnumerable<Tuple<Permutation, double>> data) { 281 Tuple<Permutation, double> prev = null, prev2 = null; 282 foreach (var d in data) { 283 if (prev == null) { 284 prev = d; 285 continue; 286 } 287 if (prev2 == null) { 288 prev2 = prev; 289 prev = d; 290 continue; 291 } 292 var dist = Dist(prev2.Item1, d.Item1); 293 yield return Tuple.Create(prev.Item1, (d.Item2 - prev2.Item2) / (double)dist); 294 prev2 = prev; 295 prev = d; 296 } 297 } 298 231 299 private static double Dist(Permutation a, Permutation b) { 232 return a.Where((t, i) => t != b[i]).Count(); 300 var dist = 0; 301 for (var i = 0; i < a.Length; i++) 302 if (a[i] != b[i]) dist++; 303 return dist; 233 304 } 234 305 235 306 private static int[] GetInverse(Permutation p) { 236 307 var inv = new int[p.Length]; 237 for (var i = 0; i < p.Length; i++) inv[p[i]] = i; 308 for (var i = 0; i < p.Length; i++) { 309 inv[p[i]] = i; 310 } 238 311 return inv; 239 312 } 240 313 241 private static void Rot1(Permutation p) { 242 var first = p[0]; 243 for (var i = 0; i < p.Length - 1; i++) p[i] = p[i + 1]; 244 p[p.Length - 1] = first; 314 // permutation must be strictly different in every position 315 private static void BiasedShuffle(Permutation p, IRandom random) { 316 for (var i = p.Length - 1; i > 0; i--) { 317 // Swap element "i" with a random earlier element (excluding itself) 318 var swapIndex = random.Next(i); 319 var h = p[swapIndex]; 320 p[swapIndex] = p[i]; 321 p[i] = h; 322 } 245 323 } 246 324 }
Note: See TracChangeset
for help on using the changeset viewer.