[17197] | 1 | using System;
|
---|
| 2 | using System.Collections.Generic;
|
---|
[17200] | 3 | using System.Diagnostics;
|
---|
[17197] | 4 | using System.Linq;
|
---|
| 5 | using System.Runtime.InteropServices;
|
---|
| 6 | using HeuristicLab.Common;
|
---|
| 7 | using HeuristicLab.Encodings.SymbolicExpressionTreeEncoding;
|
---|
| 8 |
|
---|
| 9 | namespace HeuristicLab.Problems.DataAnalysis.Symbolic.Regression.Extensions {
|
---|
| 10 | internal class ConstrainedNLSInternal {
|
---|
| 11 | private readonly int maxIterations;
|
---|
| 12 | public int MaxIterations => maxIterations;
|
---|
| 13 |
|
---|
| 14 | private readonly string solver;
|
---|
| 15 | public string Solver => solver;
|
---|
| 16 |
|
---|
| 17 | private readonly ISymbolicExpressionTree expr;
|
---|
| 18 | public ISymbolicExpressionTree Expr => expr;
|
---|
| 19 |
|
---|
| 20 | private readonly IRegressionProblemData problemData;
|
---|
| 21 |
|
---|
| 22 | public IRegressionProblemData ProblemData => problemData;
|
---|
| 23 |
|
---|
| 24 |
|
---|
| 25 | public event Action FunctionEvaluated;
|
---|
| 26 | public event Action<int, double> ConstraintEvaluated;
|
---|
| 27 |
|
---|
| 28 | private double bestError = double.MaxValue;
|
---|
| 29 | public double BestError => bestError;
|
---|
| 30 |
|
---|
| 31 | private double curError = double.MaxValue;
|
---|
| 32 | public double CurError => curError;
|
---|
| 33 |
|
---|
| 34 | private double[] bestSolution;
|
---|
| 35 | public double[] BestSolution => bestSolution;
|
---|
| 36 |
|
---|
| 37 | private ISymbolicExpressionTree bestTree;
|
---|
| 38 | public ISymbolicExpressionTree BestTree => bestTree;
|
---|
| 39 |
|
---|
| 40 | // begin internal state
|
---|
| 41 | private IntPtr nlopt;
|
---|
| 42 | private SymbolicDataAnalysisExpressionTreeLinearInterpreter interpreter;
|
---|
| 43 | private readonly NLOpt.nlopt_func calculateObjectiveDelegate; // must hold the delegate to prevent GC
|
---|
| 44 | private readonly IntPtr[] constraintDataPtr; // must hold the objects to prevent GC
|
---|
| 45 | private readonly NLOpt.nlopt_func[] calculateConstraintDelegates; // must hold the delegates to prevent GC
|
---|
| 46 | private readonly List<double> thetaValues;
|
---|
| 47 | private readonly IDictionary<string, Interval> dataIntervals;
|
---|
| 48 | private readonly int[] trainingRows;
|
---|
| 49 | private readonly double[] target;
|
---|
| 50 | private readonly ISymbolicExpressionTree preparedTree;
|
---|
| 51 | private readonly ISymbolicExpressionTreeNode[] preparedTreeParameterNodes;
|
---|
| 52 | private readonly List<ConstantTreeNode>[] allThetaNodes;
|
---|
| 53 | private readonly double[] fi_eval;
|
---|
| 54 | private readonly double[,] jac_eval;
|
---|
| 55 | private readonly ISymbolicExpressionTree scaledTree;
|
---|
| 56 |
|
---|
| 57 | // end internal state
|
---|
| 58 |
|
---|
| 59 |
|
---|
| 60 | // for data exchange to/from optimizer in native code
|
---|
| 61 | [StructLayout(LayoutKind.Sequential)]
|
---|
| 62 | private struct ConstraintData {
|
---|
| 63 | public int Idx;
|
---|
| 64 | public ISymbolicExpressionTree Tree;
|
---|
| 65 | public ISymbolicExpressionTreeNode[] ParameterNodes;
|
---|
| 66 | }
|
---|
| 67 |
|
---|
| 68 | internal ConstrainedNLSInternal(string solver, ISymbolicExpressionTree expr, int maxIterations, IRegressionProblemData problemData, double ftol_rel = 0, double ftol_abs = 0, double maxTime = 0) {
|
---|
| 69 | this.solver = solver;
|
---|
| 70 | this.expr = expr;
|
---|
| 71 | this.maxIterations = maxIterations;
|
---|
| 72 | this.problemData = problemData;
|
---|
| 73 | this.interpreter = new SymbolicDataAnalysisExpressionTreeLinearInterpreter();
|
---|
| 74 |
|
---|
| 75 |
|
---|
| 76 | var intervalConstraints = problemData.IntervalConstraints;
|
---|
| 77 | dataIntervals = problemData.VariableRanges.GetIntervals();
|
---|
| 78 | trainingRows = problemData.TrainingIndices.ToArray();
|
---|
| 79 | // buffers
|
---|
| 80 | target = problemData.TargetVariableTrainingValues.ToArray();
|
---|
| 81 | var targetStDev = target.StandardDeviationPop();
|
---|
| 82 | var targetVariance = targetStDev * targetStDev;
|
---|
| 83 | var targetMean = target.Average();
|
---|
| 84 | var pred = interpreter.GetSymbolicExpressionTreeValues(expr, problemData.Dataset, trainingRows).ToArray();
|
---|
| 85 |
|
---|
| 86 | if (pred.Any(pi => double.IsInfinity(pi) || double.IsNaN(pi))) throw new ArgumentException("The expression produces NaN or infinite values.");
|
---|
| 87 |
|
---|
| 88 | #region linear scaling
|
---|
| 89 | var predStDev = pred.StandardDeviationPop();
|
---|
| 90 | if (predStDev == 0) throw new ArgumentException("The expression is constant.");
|
---|
| 91 | var predMean = pred.Average();
|
---|
| 92 |
|
---|
| 93 | var scalingFactor = targetStDev / predStDev;
|
---|
| 94 | var offset = targetMean - predMean * scalingFactor;
|
---|
| 95 |
|
---|
| 96 | scaledTree = CopyAndScaleTree(expr, scalingFactor, offset);
|
---|
| 97 | #endregion
|
---|
| 98 |
|
---|
| 99 | // convert constants to variables named theta...
|
---|
| 100 | var treeForDerivation = ReplaceConstWithVar(scaledTree, out List<string> thetaNames, out thetaValues); // copies the tree
|
---|
| 101 |
|
---|
| 102 | // create trees for relevant derivatives
|
---|
| 103 | Dictionary<string, ISymbolicExpressionTree> derivatives = new Dictionary<string, ISymbolicExpressionTree>();
|
---|
| 104 | allThetaNodes = thetaNames.Select(_ => new List<ConstantTreeNode>()).ToArray();
|
---|
| 105 | var constraintTrees = new List<ISymbolicExpressionTree>();
|
---|
| 106 | foreach (var constraint in intervalConstraints.Constraints) {
|
---|
| 107 | if (constraint.IsDerivation) {
|
---|
| 108 | if (!problemData.AllowedInputVariables.Contains(constraint.Variable))
|
---|
| 109 | throw new ArgumentException($"Invalid constraint: the variable {constraint.Variable} does not exist in the dataset.");
|
---|
| 110 | var df = DerivativeCalculator.Derive(treeForDerivation, constraint.Variable);
|
---|
| 111 |
|
---|
| 112 | // NLOpt requires constraint expressions of the form c(x) <= 0
|
---|
| 113 | // -> we make two expressions, one for the lower bound and one for the upper bound
|
---|
| 114 |
|
---|
| 115 | if (constraint.Interval.UpperBound < double.PositiveInfinity) {
|
---|
| 116 | var df_smaller_upper = Subtract((ISymbolicExpressionTree)df.Clone(), CreateConstant(constraint.Interval.UpperBound));
|
---|
| 117 | // convert variables named theta back to constants
|
---|
| 118 | var df_prepared = ReplaceVarWithConst(df_smaller_upper, thetaNames, thetaValues, allThetaNodes);
|
---|
| 119 | constraintTrees.Add(df_prepared);
|
---|
| 120 | }
|
---|
| 121 | if (constraint.Interval.LowerBound > double.NegativeInfinity) {
|
---|
| 122 | var df_larger_lower = Subtract(CreateConstant(constraint.Interval.LowerBound), (ISymbolicExpressionTree)df.Clone());
|
---|
| 123 | // convert variables named theta back to constants
|
---|
| 124 | var df_prepared = ReplaceVarWithConst(df_larger_lower, thetaNames, thetaValues, allThetaNodes);
|
---|
| 125 | constraintTrees.Add(df_prepared);
|
---|
| 126 | }
|
---|
| 127 | } else {
|
---|
| 128 | if (constraint.Interval.UpperBound < double.PositiveInfinity) {
|
---|
| 129 | var f_smaller_upper = Subtract((ISymbolicExpressionTree)treeForDerivation.Clone(), CreateConstant(constraint.Interval.UpperBound));
|
---|
| 130 | // convert variables named theta back to constants
|
---|
| 131 | var df_prepared = ReplaceVarWithConst(f_smaller_upper, thetaNames, thetaValues, allThetaNodes);
|
---|
| 132 | constraintTrees.Add(df_prepared);
|
---|
| 133 | }
|
---|
| 134 | if (constraint.Interval.LowerBound > double.NegativeInfinity) {
|
---|
| 135 | var f_larger_lower = Subtract(CreateConstant(constraint.Interval.LowerBound), (ISymbolicExpressionTree)treeForDerivation.Clone());
|
---|
| 136 | // convert variables named theta back to constants
|
---|
| 137 | var df_prepared = ReplaceVarWithConst(f_larger_lower, thetaNames, thetaValues, allThetaNodes);
|
---|
| 138 | constraintTrees.Add(df_prepared);
|
---|
| 139 | }
|
---|
| 140 | }
|
---|
| 141 | }
|
---|
| 142 |
|
---|
| 143 | preparedTree = ReplaceVarWithConst(treeForDerivation, thetaNames, thetaValues, allThetaNodes);
|
---|
| 144 | preparedTreeParameterNodes = GetParameterNodes(preparedTree, allThetaNodes);
|
---|
| 145 |
|
---|
| 146 | var dim = thetaValues.Count;
|
---|
| 147 | fi_eval = new double[target.Length]; // init buffer;
|
---|
| 148 | jac_eval = new double[target.Length, dim]; // init buffer
|
---|
| 149 |
|
---|
| 150 |
|
---|
| 151 | var minVal = Math.Min(-1000.0, thetaValues.Min());
|
---|
| 152 | var maxVal = Math.Max(1000.0, thetaValues.Max());
|
---|
| 153 | var lb = Enumerable.Repeat(minVal, thetaValues.Count).ToArray();
|
---|
| 154 | var up = Enumerable.Repeat(maxVal, thetaValues.Count).ToArray();
|
---|
| 155 | nlopt = NLOpt.nlopt_create(GetSolver(solver), (uint)dim);
|
---|
| 156 |
|
---|
| 157 | NLOpt.nlopt_set_lower_bounds(nlopt, lb);
|
---|
| 158 | NLOpt.nlopt_set_upper_bounds(nlopt, up);
|
---|
| 159 | calculateObjectiveDelegate = new NLOpt.nlopt_func(CalculateObjective); // keep a reference to the delegate (see below)
|
---|
| 160 | NLOpt.nlopt_set_min_objective(nlopt, calculateObjectiveDelegate, IntPtr.Zero);
|
---|
| 161 |
|
---|
| 162 |
|
---|
| 163 | constraintDataPtr = new IntPtr[constraintTrees.Count];
|
---|
| 164 | calculateConstraintDelegates = new NLOpt.nlopt_func[constraintTrees.Count]; // make sure we keep a reference to the delegates (otherwise GC will free delegate objects see https://stackoverflow.com/questions/7302045/callback-delegates-being-collected#7302258)
|
---|
| 165 | for (int i = 0; i < constraintTrees.Count; i++) {
|
---|
| 166 | var constraintData = new ConstraintData() { Idx = i, Tree = constraintTrees[i], ParameterNodes = GetParameterNodes(constraintTrees[i], allThetaNodes) };
|
---|
| 167 | constraintDataPtr[i] = Marshal.AllocHGlobal(Marshal.SizeOf<ConstraintData>());
|
---|
| 168 | Marshal.StructureToPtr(constraintData, constraintDataPtr[i], fDeleteOld: false);
|
---|
| 169 | calculateConstraintDelegates[i] = new NLOpt.nlopt_func(CalculateConstraint);
|
---|
| 170 | NLOpt.nlopt_add_inequality_constraint(nlopt, calculateConstraintDelegates[i], constraintDataPtr[i], 1e-8);
|
---|
| 171 | }
|
---|
| 172 |
|
---|
| 173 | NLOpt.nlopt_set_ftol_rel(nlopt, ftol_rel);
|
---|
| 174 | NLOpt.nlopt_set_ftol_abs(nlopt, ftol_abs);
|
---|
| 175 | NLOpt.nlopt_set_maxtime(nlopt, maxTime);
|
---|
| 176 | NLOpt.nlopt_set_maxeval(nlopt, maxIterations);
|
---|
| 177 | }
|
---|
| 178 |
|
---|
| 179 | ~ConstrainedNLSInternal() {
|
---|
| 180 | if (nlopt != IntPtr.Zero)
|
---|
| 181 | NLOpt.nlopt_destroy(nlopt);
|
---|
| 182 | if (constraintDataPtr != null) {
|
---|
| 183 | for (int i = 0; i < constraintDataPtr.Length; i++)
|
---|
| 184 | if (constraintDataPtr[i] != IntPtr.Zero)
|
---|
| 185 | Marshal.FreeHGlobal(constraintDataPtr[i]);
|
---|
| 186 | }
|
---|
| 187 | }
|
---|
| 188 |
|
---|
| 189 |
|
---|
| 190 | internal void Optimize() {
|
---|
| 191 | var x = thetaValues.ToArray(); /* initial guess */
|
---|
| 192 | double minf = double.MaxValue; /* minimum objective value upon return */
|
---|
| 193 | var res = NLOpt.nlopt_optimize(nlopt, x, ref minf);
|
---|
| 194 | bestSolution = x;
|
---|
| 195 | bestError = minf;
|
---|
[17200] | 196 |
|
---|
[17197] | 197 | if (res < 0) {
|
---|
| 198 | throw new InvalidOperationException($"NLOpt failed {res} {NLOpt.nlopt_get_errmsg(nlopt)}");
|
---|
| 199 | } else {
|
---|
| 200 | // update parameters in tree
|
---|
| 201 | var pIdx = 0;
|
---|
| 202 | // here we lose the two last parameters (for linear scaling)
|
---|
| 203 | foreach (var node in scaledTree.IterateNodesPostfix()) {
|
---|
| 204 | if (node is ConstantTreeNode constTreeNode) {
|
---|
| 205 | constTreeNode.Value = x[pIdx++];
|
---|
| 206 | } else if (node is VariableTreeNode varTreeNode) {
|
---|
| 207 | varTreeNode.Weight = x[pIdx++];
|
---|
| 208 | }
|
---|
| 209 | }
|
---|
| 210 | if (pIdx != x.Length) throw new InvalidProgramException();
|
---|
| 211 | }
|
---|
| 212 | bestTree = scaledTree;
|
---|
| 213 | }
|
---|
| 214 |
|
---|
| 215 | double CalculateObjective(uint dim, double[] curX, double[] grad, IntPtr data) {
|
---|
| 216 | UpdateThetaValues(curX);
|
---|
| 217 | var sse = 0.0;
|
---|
| 218 |
|
---|
| 219 | if (grad != null) {
|
---|
| 220 | var autoDiffEval = new VectorAutoDiffEvaluator();
|
---|
| 221 | autoDiffEval.Evaluate(preparedTree, problemData.Dataset, trainingRows,
|
---|
| 222 | preparedTreeParameterNodes, fi_eval, jac_eval);
|
---|
| 223 |
|
---|
| 224 | // calc sum of squared errors and gradient
|
---|
| 225 | for (int j = 0; j < grad.Length; j++) grad[j] = 0;
|
---|
| 226 | for (int i = 0; i < target.Length; i++) {
|
---|
| 227 | var r = target[i] - fi_eval[i];
|
---|
| 228 | sse += r * r;
|
---|
| 229 | for (int j = 0; j < grad.Length; j++) {
|
---|
| 230 | grad[j] -= 2 * r * jac_eval[i, j];
|
---|
| 231 | }
|
---|
| 232 | }
|
---|
| 233 | // average
|
---|
| 234 | for (int j = 0; j < grad.Length; j++) { grad[j] /= target.Length; }
|
---|
[17200] | 235 |
|
---|
| 236 | #region check gradient
|
---|
| 237 | var eval = new VectorEvaluator();
|
---|
| 238 | for (int i = 0; i < dim; i++) {
|
---|
| 239 | // make two additional evaluations
|
---|
| 240 | var xForNumDiff = (double[])curX.Clone();
|
---|
| 241 | double delta = Math.Abs(xForNumDiff[i] * 1e-5);
|
---|
| 242 | xForNumDiff[i] += delta;
|
---|
| 243 | UpdateThetaValues(xForNumDiff);
|
---|
| 244 | var evalHigh = eval.Evaluate(preparedTree, problemData.Dataset, trainingRows);
|
---|
| 245 | var mseHigh = MSE(target, evalHigh);
|
---|
| 246 | xForNumDiff[i] = curX[i] - delta;
|
---|
| 247 | UpdateThetaValues(xForNumDiff);
|
---|
| 248 | var evalLow = eval.Evaluate(preparedTree, problemData.Dataset, trainingRows);
|
---|
| 249 | var mseLow = MSE(target, evalLow);
|
---|
| 250 |
|
---|
| 251 | var numericDiff = (mseHigh - mseLow) / (2 * delta);
|
---|
| 252 | var autoDiff = grad[i];
|
---|
| 253 | if ((Math.Abs(autoDiff) < 1e-10 && Math.Abs(numericDiff) > 1e-2)
|
---|
| 254 | || (Math.Abs(autoDiff) >= 1e-10 && Math.Abs((numericDiff - autoDiff) / numericDiff) > 1e-2))
|
---|
| 255 | throw new InvalidProgramException();
|
---|
| 256 | }
|
---|
| 257 | #endregion
|
---|
[17197] | 258 | } else {
|
---|
| 259 | var eval = new VectorEvaluator();
|
---|
| 260 | var prediction = eval.Evaluate(preparedTree, problemData.Dataset, trainingRows);
|
---|
| 261 |
|
---|
[17200] | 262 | // calc sum of squared errors
|
---|
[17197] | 263 | sse = 0.0;
|
---|
| 264 | for (int i = 0; i < target.Length; i++) {
|
---|
| 265 | var r = target[i] - prediction[i];
|
---|
| 266 | sse += r * r;
|
---|
| 267 | }
|
---|
| 268 | }
|
---|
| 269 |
|
---|
| 270 | UpdateBestSolution(sse / target.Length, curX);
|
---|
| 271 | RaiseFunctionEvaluated();
|
---|
| 272 |
|
---|
| 273 | if (double.IsNaN(sse)) return double.MaxValue;
|
---|
| 274 | return sse / target.Length;
|
---|
| 275 | }
|
---|
| 276 |
|
---|
[17200] | 277 | private double MSE(double[] a, double[] b) {
|
---|
| 278 | Trace.Assert(a.Length == b.Length);
|
---|
| 279 | var sse = 0.0;
|
---|
| 280 | for (int i = 0; i < a.Length; i++) sse += (a[i] - b[i]) * (a[i] - b[i]);
|
---|
| 281 | return sse / a.Length;
|
---|
| 282 | }
|
---|
| 283 |
|
---|
[17197] | 284 | private void UpdateBestSolution(double curF, double[] curX) {
|
---|
| 285 | if (double.IsNaN(curF) || double.IsInfinity(curF)) return;
|
---|
| 286 | else if (curF < bestError) {
|
---|
| 287 | bestError = curF;
|
---|
| 288 | bestSolution = (double[])curX.Clone();
|
---|
| 289 | }
|
---|
| 290 | }
|
---|
| 291 |
|
---|
| 292 | private void UpdateConstraintViolations(int constraintIdx, double value) {
|
---|
| 293 | if (double.IsNaN(value) || double.IsInfinity(value)) return;
|
---|
| 294 | RaiseConstraintEvaluated(constraintIdx, value);
|
---|
| 295 | // else if (curF < bestError) {
|
---|
| 296 | // bestError = curF;
|
---|
| 297 | // bestSolution = (double[])curX.Clone();
|
---|
| 298 | // }
|
---|
| 299 | }
|
---|
| 300 |
|
---|
| 301 | double CalculateConstraint(uint dim, double[] curX, double[] grad, IntPtr data) {
|
---|
| 302 | UpdateThetaValues(curX);
|
---|
| 303 | var intervalEvaluator = new IntervalEvaluator();
|
---|
| 304 | var constraintData = Marshal.PtrToStructure<ConstraintData>(data);
|
---|
| 305 |
|
---|
| 306 | if (grad != null) for (int j = 0; j < grad.Length; j++) grad[j] = 0; // clear grad
|
---|
| 307 |
|
---|
| 308 | var interval = intervalEvaluator.Evaluate(constraintData.Tree, dataIntervals, constraintData.ParameterNodes,
|
---|
| 309 | out double[] lowerGradient, out double[] upperGradient);
|
---|
| 310 |
|
---|
| 311 | // we transformed this to a constraint c(x) <= 0, so only the upper bound is relevant for us
|
---|
| 312 | if (grad != null) for (int j = 0; j < grad.Length; j++) { grad[j] = upperGradient[j]; }
|
---|
[17200] | 313 |
|
---|
| 314 | #region check gradient
|
---|
| 315 | for (int i = 0; i < dim; i++) {
|
---|
| 316 | // make two additional evaluations
|
---|
| 317 | var xForNumDiff = (double[])curX.Clone();
|
---|
| 318 | double delta = Math.Abs(xForNumDiff[i] * 1e-5);
|
---|
| 319 | xForNumDiff[i] += delta;
|
---|
| 320 | UpdateThetaValues(xForNumDiff);
|
---|
| 321 | var evalHigh = intervalEvaluator.Evaluate(constraintData.Tree, dataIntervals, constraintData.ParameterNodes,
|
---|
| 322 | out double[] unusedLowerGradientHigh, out double[] unusedUpperGradientHigh);
|
---|
| 323 |
|
---|
| 324 | xForNumDiff[i] = curX[i] - delta;
|
---|
| 325 | UpdateThetaValues(xForNumDiff);
|
---|
| 326 | var evalLow = intervalEvaluator.Evaluate(constraintData.Tree, dataIntervals, constraintData.ParameterNodes,
|
---|
| 327 | out double[] unusedLowerGradientLow, out double[] unusedUpperGradientLow);
|
---|
| 328 |
|
---|
| 329 | var numericDiff = (evalHigh.UpperBound - evalLow.UpperBound) / (2 * delta);
|
---|
| 330 | var autoDiff = grad[i];
|
---|
| 331 | // XXX GRADIENT FOR UPPER BOUND FOR FUNCTION IS ZERO FOR THE FIRST SET OF VARIABLES? GRADIENT FOR ADDITION INCORRECT?!
|
---|
| 332 | if ((Math.Abs(autoDiff) < 1e-10 && Math.Abs(numericDiff) > 1e-2)
|
---|
| 333 | || (Math.Abs(autoDiff) >= 1e-10 && Math.Abs((numericDiff - autoDiff) / numericDiff) > 1e-2))
|
---|
| 334 | throw new InvalidProgramException();
|
---|
| 335 | }
|
---|
| 336 | #endregion
|
---|
| 337 |
|
---|
| 338 |
|
---|
[17197] | 339 | UpdateConstraintViolations(constraintData.Idx, interval.UpperBound);
|
---|
| 340 | if (double.IsNaN(interval.UpperBound)) return double.MaxValue;
|
---|
| 341 | else return interval.UpperBound;
|
---|
| 342 | }
|
---|
| 343 |
|
---|
| 344 |
|
---|
| 345 | void UpdateThetaValues(double[] theta) {
|
---|
| 346 | for (int i = 0; i < theta.Length; ++i) {
|
---|
| 347 | foreach (var constNode in allThetaNodes[i]) constNode.Value = theta[i];
|
---|
| 348 | }
|
---|
| 349 | }
|
---|
| 350 |
|
---|
| 351 | internal void RequestStop() {
|
---|
| 352 | NLOpt.nlopt_set_force_stop(nlopt, 1); // hopefully NLOpt is thread safe , val must be <> 0 otherwise no effect
|
---|
| 353 | }
|
---|
| 354 |
|
---|
| 355 | private void RaiseFunctionEvaluated() {
|
---|
| 356 | FunctionEvaluated?.Invoke();
|
---|
| 357 | }
|
---|
| 358 |
|
---|
| 359 | private void RaiseConstraintEvaluated(int idx, double value) {
|
---|
| 360 | ConstraintEvaluated?.Invoke(idx, value);
|
---|
| 361 | }
|
---|
| 362 |
|
---|
| 363 |
|
---|
| 364 | #region helper
|
---|
| 365 |
|
---|
| 366 | private static ISymbolicExpressionTree CopyAndScaleTree(ISymbolicExpressionTree tree, double scalingFactor, double offset) {
|
---|
| 367 | var m = (ISymbolicExpressionTree)tree.Clone();
|
---|
| 368 |
|
---|
| 369 | var add = MakeNode<Addition>(MakeNode<Multiplication>(m.Root.GetSubtree(0).GetSubtree(0), CreateConstant(scalingFactor)), CreateConstant(offset));
|
---|
| 370 | m.Root.GetSubtree(0).RemoveSubtree(0);
|
---|
| 371 | m.Root.GetSubtree(0).AddSubtree(add);
|
---|
| 372 | return m;
|
---|
| 373 | }
|
---|
| 374 |
|
---|
| 375 | private static void UpdateConstants(ISymbolicExpressionTreeNode[] nodes, double[] constants) {
|
---|
| 376 | if (nodes.Length != constants.Length) throw new InvalidOperationException();
|
---|
| 377 | for (int i = 0; i < nodes.Length; i++) {
|
---|
| 378 | if (nodes[i] is VariableTreeNode varNode) varNode.Weight = constants[i];
|
---|
| 379 | else if (nodes[i] is ConstantTreeNode constNode) constNode.Value = constants[i];
|
---|
| 380 | }
|
---|
| 381 | }
|
---|
| 382 |
|
---|
| 383 | private NLOpt.nlopt_algorithm GetSolver(string solver) {
|
---|
| 384 | if (solver.Contains("MMA")) return NLOpt.nlopt_algorithm.NLOPT_LD_MMA;
|
---|
| 385 | if (solver.Contains("COBYLA")) return NLOpt.nlopt_algorithm.NLOPT_LN_COBYLA;
|
---|
| 386 | if (solver.Contains("CCSAQ")) return NLOpt.nlopt_algorithm.NLOPT_LD_CCSAQ;
|
---|
| 387 | if (solver.Contains("ISRES")) return NLOpt.nlopt_algorithm.NLOPT_GN_ISRES;
|
---|
| 388 | throw new ArgumentException($"Unknown solver {solver}");
|
---|
| 389 | }
|
---|
| 390 |
|
---|
| 391 | private static ISymbolicExpressionTreeNode[] GetParameterNodes(ISymbolicExpressionTree tree, List<ConstantTreeNode>[] allNodes) {
|
---|
| 392 | // TODO better solution necessary
|
---|
| 393 | var treeConstNodes = tree.IterateNodesPostfix().OfType<ConstantTreeNode>().ToArray();
|
---|
| 394 | var paramNodes = new ISymbolicExpressionTreeNode[allNodes.Length];
|
---|
| 395 | for (int i = 0; i < paramNodes.Length; i++) {
|
---|
| 396 | paramNodes[i] = allNodes[i].SingleOrDefault(n => treeConstNodes.Contains(n));
|
---|
| 397 | }
|
---|
| 398 | return paramNodes;
|
---|
| 399 | }
|
---|
| 400 |
|
---|
| 401 | private static ISymbolicExpressionTree ReplaceVarWithConst(ISymbolicExpressionTree tree, List<string> thetaNames, List<double> thetaValues, List<ConstantTreeNode>[] thetaNodes) {
|
---|
| 402 | var copy = (ISymbolicExpressionTree)tree.Clone();
|
---|
| 403 | var nodes = copy.IterateNodesPostfix().ToList();
|
---|
| 404 | for (int i = 0; i < nodes.Count; i++) {
|
---|
| 405 | var n = nodes[i] as VariableTreeNode;
|
---|
| 406 | if (n != null) {
|
---|
| 407 | var thetaIdx = thetaNames.IndexOf(n.VariableName);
|
---|
| 408 | if (thetaIdx >= 0) {
|
---|
| 409 | var parent = n.Parent;
|
---|
| 410 | if (thetaNodes[thetaIdx].Any()) {
|
---|
| 411 | // HACK: REUSE CONSTANT TREE NODE IN SEVERAL TREES
|
---|
| 412 | // we use this trick to allow autodiff over thetas when thetas occurr multiple times in the tree (e.g. in derived trees)
|
---|
| 413 | var constNode = thetaNodes[thetaIdx].First();
|
---|
| 414 | var childIdx = parent.IndexOfSubtree(n);
|
---|
| 415 | parent.RemoveSubtree(childIdx);
|
---|
| 416 | parent.InsertSubtree(childIdx, constNode);
|
---|
| 417 | } else {
|
---|
| 418 | var constNode = (ConstantTreeNode)CreateConstant(thetaValues[thetaIdx]);
|
---|
| 419 | var childIdx = parent.IndexOfSubtree(n);
|
---|
| 420 | parent.RemoveSubtree(childIdx);
|
---|
| 421 | parent.InsertSubtree(childIdx, constNode);
|
---|
| 422 | thetaNodes[thetaIdx].Add(constNode);
|
---|
| 423 | }
|
---|
| 424 | }
|
---|
| 425 | }
|
---|
| 426 | }
|
---|
| 427 | return copy;
|
---|
| 428 | }
|
---|
| 429 |
|
---|
| 430 | private static ISymbolicExpressionTree ReplaceConstWithVar(ISymbolicExpressionTree tree, out List<string> thetaNames, out List<double> thetaValues) {
|
---|
| 431 | thetaNames = new List<string>();
|
---|
| 432 | thetaValues = new List<double>();
|
---|
| 433 | var copy = (ISymbolicExpressionTree)tree.Clone();
|
---|
| 434 | var nodes = copy.IterateNodesPostfix().ToList();
|
---|
| 435 |
|
---|
| 436 | int n = 1;
|
---|
| 437 | for (int i = 0; i < nodes.Count; ++i) {
|
---|
| 438 | var node = nodes[i];
|
---|
| 439 | if (node is ConstantTreeNode constantTreeNode) {
|
---|
| 440 | var thetaVar = (VariableTreeNode)new Problems.DataAnalysis.Symbolic.Variable().CreateTreeNode();
|
---|
| 441 | thetaVar.Weight = 1;
|
---|
| 442 | thetaVar.VariableName = $"θ{n++}";
|
---|
| 443 |
|
---|
| 444 | thetaNames.Add(thetaVar.VariableName);
|
---|
| 445 | thetaValues.Add(constantTreeNode.Value);
|
---|
| 446 |
|
---|
| 447 | var parent = constantTreeNode.Parent;
|
---|
| 448 | if (parent != null) {
|
---|
| 449 | var index = constantTreeNode.Parent.IndexOfSubtree(constantTreeNode);
|
---|
| 450 | parent.RemoveSubtree(index);
|
---|
| 451 | parent.InsertSubtree(index, thetaVar);
|
---|
| 452 | }
|
---|
| 453 | }
|
---|
| 454 | if (node is VariableTreeNode varTreeNode) {
|
---|
| 455 | var thetaVar = (VariableTreeNode)new Problems.DataAnalysis.Symbolic.Variable().CreateTreeNode();
|
---|
| 456 | thetaVar.Weight = 1;
|
---|
| 457 | thetaVar.VariableName = $"θ{n++}";
|
---|
| 458 |
|
---|
| 459 | thetaNames.Add(thetaVar.VariableName);
|
---|
| 460 | thetaValues.Add(varTreeNode.Weight);
|
---|
| 461 |
|
---|
| 462 | var parent = varTreeNode.Parent;
|
---|
| 463 | if (parent != null) {
|
---|
| 464 | var index = varTreeNode.Parent.IndexOfSubtree(varTreeNode);
|
---|
| 465 | parent.RemoveSubtree(index);
|
---|
| 466 | var prodNode = MakeNode<Multiplication>();
|
---|
| 467 | varTreeNode.Weight = 1.0;
|
---|
| 468 | prodNode.AddSubtree(varTreeNode);
|
---|
| 469 | prodNode.AddSubtree(thetaVar);
|
---|
| 470 | parent.InsertSubtree(index, prodNode);
|
---|
| 471 | }
|
---|
| 472 | }
|
---|
| 473 | }
|
---|
| 474 | return copy;
|
---|
| 475 | }
|
---|
| 476 |
|
---|
| 477 | private static ISymbolicExpressionTreeNode CreateConstant(double value) {
|
---|
| 478 | var constantNode = (ConstantTreeNode)new Constant().CreateTreeNode();
|
---|
| 479 | constantNode.Value = value;
|
---|
| 480 | return constantNode;
|
---|
| 481 | }
|
---|
| 482 |
|
---|
| 483 | private static ISymbolicExpressionTree Subtract(ISymbolicExpressionTree t, ISymbolicExpressionTreeNode b) {
|
---|
| 484 | var sub = MakeNode<Subtraction>(t.Root.GetSubtree(0).GetSubtree(0), b);
|
---|
| 485 | t.Root.GetSubtree(0).RemoveSubtree(0);
|
---|
| 486 | t.Root.GetSubtree(0).InsertSubtree(0, sub);
|
---|
| 487 | return t;
|
---|
| 488 | }
|
---|
| 489 | private static ISymbolicExpressionTree Subtract(ISymbolicExpressionTreeNode b, ISymbolicExpressionTree t) {
|
---|
| 490 | var sub = MakeNode<Subtraction>(b, t.Root.GetSubtree(0).GetSubtree(0));
|
---|
| 491 | t.Root.GetSubtree(0).RemoveSubtree(0);
|
---|
| 492 | t.Root.GetSubtree(0).InsertSubtree(0, sub);
|
---|
| 493 | return t;
|
---|
| 494 | }
|
---|
| 495 |
|
---|
| 496 | private static ISymbolicExpressionTreeNode MakeNode<T>(params ISymbolicExpressionTreeNode[] fs) where T : ISymbol, new() {
|
---|
| 497 | var node = new T().CreateTreeNode();
|
---|
| 498 | foreach (var f in fs) node.AddSubtree(f);
|
---|
| 499 | return node;
|
---|
| 500 | }
|
---|
| 501 | #endregion
|
---|
| 502 | }
|
---|
| 503 | } |
---|