Fixed up the docs a bit, and changes fixed an issue in the optimizer.

Author: Thomas L. Redman

lbjava/src/main/java/edu/illinois/cs/cogcomp/lbjava/learn/featurepruning/SparseNetworkOptimizer.java

@@ -67,7 +67,7 @@ protected boolean hasWeight(Lexicon lex, Feature f){
 sum += Math.abs(avg);
 
 // if the value is sufficiently large, then we have a good weight and should keep.
-if (sum > this.threshold)
+if (sum >=this.threshold)
 returntrue;
  }
  } else{
@@ -77,7 +77,7 @@ protected boolean hasWeight(Lexicon lex, Feature f){
 sum += Math.abs(wt);
 
 // if the value is sufficiently large, then we have a good weight and should keep.
-if (sum > this.threshold)
+if (sum >=this.threshold)
 returntrue;
  }
  }
@@ -118,6 +118,7 @@ protected int[] identifyUselessFeatures(){
  } else{
 defunct.add(fi);
  }
+
 this.uselessFeatureNames.add(entry.getKey());
 count++;
  }

lbjava/src/main/java/edu/illinois/cs/cogcomp/lbjava/learn/featurepruning/package-info.java

@@ -6,27 +6,33 @@
  * http://cogcomp.cs.illinois.edu/
  */
 /**
- * <p>
- * For sparse learners, it is often the case that the array of features you learn 
+ * <p>For sparse learners, it is often the case that the array of features you learn 
  * contains only a subset of useful features. When we leave these features in the lexicon,
  * we end up with bloated lexicons and weight vectors. This leads to larger than necessary
- * models.<p>
+ * models.</p>
  * 
- * This package contains an interface that defines the life cycle for the pruning
+ * <p>This package contains an interface that defines the life cycle for the feature pruning
  * process, as well as some implementations, one that takes multiple weight vectors (for
- * multi-class network learners), and some that takes only one weight vector.<p>
+ * multi-class network learners), and some that takes only one weight vector.</p>
  * 
- * All optimizers should subclass @see LexiconOptimizer which implements most of the
+ * <p>All optimizers should subclass @see LexiconOptimizer which implements most of the
  * optimization. Subclass will need to provide methods to compute the weight value to compare
  * against the threshold, a method to identify the useless features, and a method to prune
- * those features.<p>
+ * those features.</p>
  * 
- * The optimizers are invoked by the doneTraining method of the Learner class when all learning
- * is complete. For those who have build their own training procedure, they are required to invoke 
- * the doneTraining and startTraining method during their training process.<p>
+ * <p>The optimizers are invoked by the{@link edu.illinois.cs.cogcomp.lbjava.learn.Learner#doneTraining}
+ * method of the Learner class when all learning is complete. For those learners that include a feature
+ * pruning implementation, they must override this method to invoke the optimizer. In this way, during the
+ * normal LBJava compile and model build cycle, the optimization is performed automatically. For those
+ * who have build their own training procedure, they are required to invoke the doneTraining and 
+ *{@link edu.illinois.cs.cogcomp.lbjava.learn.Learner#startTraining} method at appropriate points during
+ * their training process.</p>
  * 
- * The pruning threshold value is provided by the specific learner, and should be, in one way or 
- * another, parameterized.<p>
+ * <p>The learner classes typically have a parameter that can be set to change the default feature
+ * pruning threshold to any the user might choose, or it can be set to 0.0 to disable. </p>
+ * 
+ * <p>The pruning threshold value is provided by the specific learner, and should be, in one way or 
+ * another, parameterized.</p>
  * @author redman
  */
 packageedu.illinois.cs.cogcomp.lbjava.learn.featurepruning;