Changeset 12014 for branches/HeuristicLab.Problems.GrammaticalOptimization/HeuristicLab.Problems.GrammaticalOptimization.SymbReg

Timestamp:

02/16/15 09:14:38 (10 years ago)

Author:

gkronber

Message:

#2283: worked on transformation of arithmetic expressions to canonical form

Location:

branches/HeuristicLab.Problems.GrammaticalOptimization/HeuristicLab.Problems.GrammaticalOptimization.SymbReg

Files:

• ExpressionCompiler.cs (modified) (3 diffs)
• SymbolicRegressionProblem.cs (modified) (3 diffs)

branches/HeuristicLab.Problems.GrammaticalOptimization/HeuristicLab.Problems.GrammaticalOptimization.SymbReg/ExpressionCompiler.cs

@@ -17 +17 @@
     // does not compile to IL it is only necessary to calculate gradients for parameter optimization
     // Expr -> Term { ('+' | '-' | '^' ) Term } 
-    // Term -> Fact { ('*' | '%') Fact }
-    // Fact -> '!' Expr | '(' Expr ')' | Var | const
+    // Term -> Fact { ('*' | '%' | '/') Fact }
+    // Fact -> '!' Expr | '(' Expr ')' | Var | const | one
     // Var -> 'a'..'z'
     // const -> '0' .. '9'
+    // one -> |    // pipe symbol for constant one instead of ERC 1
 
     // constants are completely ignored, instead we introduce a multiplicative constant factor for each term and an additive constant term for each expression
@@ -94 +95 @@
       if (f != null) factors.Add(f);
       var curSy = CurSy();
-      while (curSy == '*' || curSy == '%') {
+      while (curSy == '*' || curSy == '%' || curSy == '/') { // division and protected division symbols are handled in the same way
         if (curSy == '*') {
           NewSy();
@@ -134 +135 @@
         r = variables[varIdx];
         NewSy();
+      } else if (curSy == '|') {
+        // pipe symbol is used in the expressionextender to represent constant one (|/x). 
+        // this is necessary because we also use symbols 0..9 as indices for ERCs
+        r = 1.0;
+        NewSy();
       } else if (curSy >= '0' && curSy <= '9') {
         int o = (byte)curSy - (byte)'0';

branches/HeuristicLab.Problems.GrammaticalOptimization/HeuristicLab.Problems.GrammaticalOptimization.SymbReg/SymbolicRegressionProblem.cs

@@ -9 +9 @@
 using System.Text;
 using AutoDiff;
+using HeuristicLab.Algorithms.Bandits;
 using HeuristicLab.Common;
 using HeuristicLab.Encodings.SymbolicExpressionTreeEncoding;
@@ -128 +129 @@
         return OptimizeConstantsAndEvaluate(sentence);
       else {
+        var extender = new ExpressionExtender();
+
+        Debug.Assert(SimpleEvaluate(sentence) == SimpleEvaluate(extender.CanonicalRepresentation(sentence)));
         return SimpleEvaluate(sentence);
       }
@@ -143 +147 @@
 
     public string CanonicalRepresentation(string phrase) {
-      return phrase;
+      var extender = new ExpressionExtender();
+      return extender.CanonicalRepresentation(phrase);
     }
 
     public IEnumerable<Feature> GetFeatures(string phrase) {
-      // first generate canonical expression (which must not contain ())
-      // recursively split into expressions
-      // for each expression split into terms
-      // for each term order factors to canonical // ../E = * 1/E
+      phrase = CanonicalRepresentation(phrase);
       return new Feature[] { new Feature(phrase, 1.0) };
     }