Changeset 16691 for branches

Timestamp:

03/16/19 13:13:39 (6 years ago)

Author:

gkronber

Message:

#2974: improved code for benchmarking (previous code failed to account for update of trees after call of Optimize as well as early stopping in Optimize for trees with invalid outputs)

File:

• branches/2974_Constants_Optimization/Benchmarks/Benchmarks/ConstantsOptimizationPerformance.cs (modified) (3 diffs)

branches/2974_Constants_Optimization/Benchmarks/Benchmarks/ConstantsOptimizationPerformance.cs

@@ -16 +16 @@
     private static readonly int seed = 1234;
     private static readonly int maxTreeSize = 50;
-    private static readonly int nTrees = 1000;
+    private static readonly int nTrees = 10;
 
-    private ISymbolicExpressionTree[] trees;
+    private List<ISymbolicExpressionTree> globalTrees = new List<ISymbolicExpressionTree>();
+    private List<ISymbolicExpressionTree> iterationTrees = new List<ISymbolicExpressionTree>();
     private IRegressionProblemData problemData;
     private TypeCoherentExpressionGrammar grammar;
@@ -34 +35 @@
       grammar.ConfigureAsDefaultRegressionGrammar();
 
-      trees = Util.CreateRandomTrees(twister, problemData.Dataset, grammar, nTrees, 1, maxTreeSize, 0, 0);
-      foreach (SymbolicExpressionTree tree in trees) {
-        Util.InitTree(tree, twister, problemData.AllowedInputVariables.ToList());
+      // generate random trees and only keep trees which are ok
+      while (globalTrees.Count < nTrees) {
+        var t = Util.CreateRandomTrees(twister, problemData.Dataset, grammar, 1, 1, maxTreeSize, 0, 0).Single();
+        Util.InitTree(t, twister, problemData.AllowedInputVariables.ToList());
+        var q = SymbolicRegressionSingleObjectivePearsonRSquaredEvaluator.Calculate(interpreter, t, double.MinValue, double.MaxValue, problemData, rows, applyLinearScaling: true);
+        if (q > 0) globalTrees.Add(t);
+      }
+    }
+
+    [IterationSetup]
+    public void IterationSetup() {
+      iterationTrees.Clear();
+      for (int i = 0; i < globalTrees.Count; i++) {
+        var t = (ISymbolicExpressionTree)globalTrees[i].Clone(); // trees are changed in OptimizeConstants, so we need to clone the original trees for each iteration
+        iterationTrees.Add(t);
       }
     }
@@ -57 +70 @@
 
       if (Iterations > 0) {
-        for (int i = 0; i < trees.Length; i++) {
-          avgQuality += LMConstantsOptimizer.OptimizeConstants(trees[i], problemData.Dataset, problemData.TargetVariable, rows, true, Iterations);
+        for (int i = 0; i < iterationTrees.Count; i++) {
+          avgQuality += LMConstantsOptimizer.OptimizeConstants(iterationTrees[i], problemData.Dataset, problemData.TargetVariable, rows, applyLinearScaling: true, maxIterations: Iterations);
         }
       } else {
-        for (int i = 0; i < trees.Length; i++) {
-          avgQuality += SymbolicRegressionSingleObjectivePearsonRSquaredEvaluator.Calculate(interpreter, trees[i], double.MinValue, double.MaxValue, problemData, rows, true);
+        for (int i = 0; i < iterationTrees.Count; i++) {
+          avgQuality += SymbolicRegressionSingleObjectivePearsonRSquaredEvaluator.Calculate(interpreter, iterationTrees[i], double.MinValue, double.MaxValue, problemData, rows, applyLinearScaling: true);
         }
       }
-      return avgQuality / trees.Length;
+      return avgQuality / iterationTrees.Count;
     }
   }