Menu
▾
▴
[DL-Learner SVN] SF.net SVN: dl-learner:[1516] trunk
|
From: <jen...@us...> - 2008-11-15 08:43:57
|
Revision: 1516
http://dl-learner.svn.sourceforge.net/dl-learner/?rev=1516&view=rev
Author: jenslehmann
Date: 2008-11-15 08:43:53 +0000 (Sat, 15 Nov 2008)
Log Message:
-----------
added two more parameters to learning algorithm
Modified Paths:
--------------
trunk/examples/carcinogenesis/train.conf
trunk/src/dl-learner/org/dllearner/algorithms/refexamples/ExampleBasedROLComponent.java
trunk/src/dl-learner/org/dllearner/algorithms/refexamples/ExampleBasedROLearner.java
trunk/src/dl-learner/org/dllearner/algorithms/refexamples/MultiHeuristic.java
trunk/src/dl-learner/org/dllearner/core/configurators/ExampleBasedROLComponentConfigurator.java
Modified: trunk/examples/carcinogenesis/train.conf
===================================================================
--- trunk/examples/carcinogenesis/train.conf 2008-11-14 16:55:52 UTC (rev 1515)
+++ trunk/examples/carcinogenesis/train.conf 2008-11-15 08:43:53 UTC (rev 1516)
@@ -1,12 +1,15 @@
import("carcinogenesis.owl");
+// store some settings to make the experiment reproducable in the future
reasoner = fastInstanceChecker;
algorithm = refexamples;
-// refexamples.noisePercentage = 28;
refexamples.noisePercentage = 32;
refexamples.startClass = "http://dl-learner.org/carcinogenesis#Compound";
refexamples.writeSearchTree = false;
refexamples.searchTreeFile = "log/carcinogenesis/searchTree.log";
+refexamples.negativeWeight = 0.8;
+refexamples.startNodeBonus = 2.0;
+refexamples.forceRefinementLengthIncrease = false;
+"http://dl-learner.org/carcinogenesis#d1"
+"http://dl-learner.org/carcinogenesis#d10"
Modified: trunk/src/dl-learner/org/dllearner/algorithms/refexamples/ExampleBasedROLComponent.java
===================================================================
--- trunk/src/dl-learner/org/dllearner/algorithms/refexamples/ExampleBasedROLComponent.java 2008-11-14 16:55:52 UTC (rev 1515)
+++ trunk/src/dl-learner/org/dllearner/algorithms/refexamples/ExampleBasedROLComponent.java 2008-11-15 08:43:53 UTC (rev 1516)
@@ -213,6 +213,8 @@
options.add(noisePercentage);
options.add(new StringConfigOption("startClass", "the named class which should be used to start the algorithm (GUI: needs a widget for selecting a class)"));
options.add(new BooleanConfigOption("forceRefinementLengthIncrease", "specifies whether nodes should be expanded until only longer refinements are reached"));
+ options.add(new DoubleConfigOption("negativeWeight", "Used to penalise errors on negative examples different from those of positive examples (lower = less importance for negatives).",1.0));
+ options.add(new DoubleConfigOption("startNodeBonus", "You can use this to give a heuristic bonus on the start node (= initially broader exploration of search space).",0.0));
return options;
}
@@ -328,9 +330,9 @@
} else {
if(learningProblem instanceof PosOnlyLP) {
// throw new RuntimeException("does not work with positive examples only yet");
- algHeuristic = new MultiHeuristic(((PosOnlyLP)learningProblem).getPositiveExamples().size(),0);
+ algHeuristic = new MultiHeuristic(((PosOnlyLP)learningProblem).getPositiveExamples().size(),0, configurator);
} else {
- algHeuristic = new MultiHeuristic(((PosNegLP)learningProblem).getPositiveExamples().size(),((PosNegLP)learningProblem).getNegativeExamples().size());
+ algHeuristic = new MultiHeuristic(((PosNegLP)learningProblem).getPositiveExamples().size(),((PosNegLP)learningProblem).getNegativeExamples().size(), configurator);
}
}
Modified: trunk/src/dl-learner/org/dllearner/algorithms/refexamples/ExampleBasedROLearner.java
===================================================================
--- trunk/src/dl-learner/org/dllearner/algorithms/refexamples/ExampleBasedROLearner.java 2008-11-14 16:55:52 UTC (rev 1515)
+++ trunk/src/dl-learner/org/dllearner/algorithms/refexamples/ExampleBasedROLearner.java 2008-11-15 08:43:53 UTC (rev 1516)
@@ -405,9 +405,9 @@
long lastPrintTime = 0;
long lastTreeTraversalTime = System.nanoTime();
long lastReductionTime = System.nanoTime();
- // try a traversal after 100 seconds
+ // try a traversal after x seconds
long traversalInterval = 300l * 1000000000l;
- long reductionInterval = 500l * 1000000000l;
+ long reductionInterval = 300l * 1000000000l;
long currentTime;
while (!solutionFound && !stop) {
Modified: trunk/src/dl-learner/org/dllearner/algorithms/refexamples/MultiHeuristic.java
===================================================================
--- trunk/src/dl-learner/org/dllearner/algorithms/refexamples/MultiHeuristic.java 2008-11-14 16:55:52 UTC (rev 1515)
+++ trunk/src/dl-learner/org/dllearner/algorithms/refexamples/MultiHeuristic.java 2008-11-15 08:43:53 UTC (rev 1516)
@@ -21,6 +21,7 @@
import java.util.List;
+import org.dllearner.core.configurators.ExampleBasedROLComponentConfigurator;
import org.dllearner.core.owl.DatatypeSomeRestriction;
import org.dllearner.core.owl.Description;
import org.dllearner.core.owl.Thing;
@@ -70,8 +71,8 @@
private ConceptComparator conceptComparator = new ConceptComparator();
// heuristic parameters
- private double expansionPenaltyFactor;
- private double gainBonusFactor;
+ private double expansionPenaltyFactor = 0.02;
+ private double gainBonusFactor = 0.5;
private double nodeChildPenalty = 0.0001; // (use higher values than 0.0001 for simple learning problems);
private double startNodeBonus = 0.1; //was 2.0
// penalise errors on positive examples harder than on negative examples
@@ -83,16 +84,28 @@
private int nrOfExamples;
public MultiHeuristic(int nrOfPositiveExamples, int nrOfNegativeExamples) {
- this(nrOfPositiveExamples, nrOfNegativeExamples, 0.02, 0.5);
+ this.nrOfNegativeExamples = nrOfNegativeExamples;
+ nrOfExamples = nrOfPositiveExamples + nrOfNegativeExamples;
+// this(nrOfPositiveExamples, nrOfNegativeExamples, 0.02, 0.5);
}
- public MultiHeuristic(int nrOfPositiveExamples, int nrOfNegativeExamples, double expansionPenaltyFactor, double gainBonusFactor) {
+ public MultiHeuristic(int nrOfPositiveExamples, int nrOfNegativeExamples, ExampleBasedROLComponentConfigurator configurator) {
this.nrOfNegativeExamples = nrOfNegativeExamples;
nrOfExamples = nrOfPositiveExamples + nrOfNegativeExamples;
- this.expansionPenaltyFactor = expansionPenaltyFactor;
- this.gainBonusFactor = gainBonusFactor;
+ negativeWeight = configurator.getNegativeWeight();
+ startNodeBonus = configurator.getStartNodeBonus();
+ System.out.println(negativeWeight);
+ System.out.println(startNodeBonus);
}
+// public MultiHeuristic(int nrOfPositiveExamples, int nrOfNegativeExamples, double expansionPenaltyFactor, double gainBonusFactor) {
+// this.nrOfNegativeExamples = nrOfNegativeExamples;
+// nrOfExamples = nrOfPositiveExamples + nrOfNegativeExamples;
+// this.expansionPenaltyFactor = expansionPenaltyFactor;
+// this.gainBonusFactor = gainBonusFactor;
+// }
+
+
/* (non-Javadoc)
* @see java.util.Comparator#compare(java.lang.Object, java.lang.Object)
*/
@@ -141,7 +154,7 @@
// do not count TOP symbols (in particular in ALL r.TOP and EXISTS r.TOP)
// as they provide no extra information
if(description instanceof Thing)
- bonus = 2;
+ bonus = 1; //2;
// if(description instanceof BooleanValueRestriction)
// bonus = -1;
@@ -149,7 +162,7 @@
// some bonus for doubles because they are already penalised by length 3
if(description instanceof DatatypeSomeRestriction) {
// System.out.println(description);
- bonus = 2;
+ bonus = 3; //2;
}
List<Description> children = description.getChildren();
Modified: trunk/src/dl-learner/org/dllearner/core/configurators/ExampleBasedROLComponentConfigurator.java
===================================================================
--- trunk/src/dl-learner/org/dllearner/core/configurators/ExampleBasedROLComponentConfigurator.java 2008-11-14 16:55:52 UTC (rev 1515)
+++ trunk/src/dl-learner/org/dllearner/core/configurators/ExampleBasedROLComponentConfigurator.java 2008-11-15 08:43:53 UTC (rev 1516)
@@ -357,6 +357,24 @@
public boolean getForceRefinementLengthIncrease() {
return (Boolean) ComponentManager.getInstance().getConfigOptionValue(exampleBasedROLComponent, "forceRefinementLengthIncrease") ;
}
+/**
+* negativeWeight Used to penalise errors on negative examples different from those of positive examples (lower = less importance for negatives)..
+* mandatory: false| reinit necessary: true
+* default value: 1.0
+* @return double
+**/
+public double getNegativeWeight() {
+return (Double) ComponentManager.getInstance().getConfigOptionValue(exampleBasedROLComponent, "negativeWeight") ;
+}
+/**
+* startNodeBonus You can use this to give a heuristic bonus on the start node (= initially broader exploration of search space)..
+* mandatory: false| reinit necessary: true
+* default value: 0.0
+* @return double
+**/
+public double getStartNodeBonus() {
+return (Double) ComponentManager.getInstance().getConfigOptionValue(exampleBasedROLComponent, "startNodeBonus") ;
+}
/**
* @param writeSearchTree specifies whether to write a search tree.
@@ -655,6 +673,24 @@
ComponentManager.getInstance().applyConfigEntry(exampleBasedROLComponent, "forceRefinementLengthIncrease", forceRefinementLengthIncrease);
reinitNecessary = true;
}
+/**
+* @param negativeWeight Used to penalise errors on negative examples different from those of positive examples (lower = less importance for negatives)..
+* mandatory: false| reinit necessary: true
+* default value: 1.0
+**/
+public void setNegativeWeight(double negativeWeight) {
+ComponentManager.getInstance().applyConfigEntry(exampleBasedROLComponent, "negativeWeight", negativeWeight);
+reinitNecessary = true;
+}
+/**
+* @param startNodeBonus You can use this to give a heuristic bonus on the start node (= initially broader exploration of search space)..
+* mandatory: false| reinit necessary: true
+* default value: 0.0
+**/
+public void setStartNodeBonus(double startNodeBonus) {
+ComponentManager.getInstance().applyConfigEntry(exampleBasedROLComponent, "startNodeBonus", startNodeBonus);
+reinitNecessary = true;
+}
/**
* true, if this component needs reinitializsation.
This was sent by the SourceForge.net collaborative development platform, the world's largest Open Source development site.
|
