svn commit: r999145 - /mahout/trunk/core/src/main/java/org/apache/mahout/classifier/sgd/AdaptiveLogisticRegression.java

[td...@apache.org]

Author: tdunning
Date: Mon Sep 20 22:26:07 2010
New Revision: 999145

URL: http://svn.apache.org/viewvc?rev=999145&view=rev
Log:
Freeze survivors of evolution deal with early good results.

Modified:
    mahout/trunk/core/src/main/java/org/apache/mahout/classifier/sgd/AdaptiveLogisticRegression.java

Modified: mahout/trunk/core/src/main/java/org/apache/mahout/classifier/sgd/AdaptiveLogisticRegression.java
URL: http://svn.apache.org/viewvc/mahout/trunk/core/src/main/java/org/apache/mahout/classifier/sgd/AdaptiveLogisticRegression.java?rev=999145&r1=999144&r2=999145&view=diff
==============================================================================
--- mahout/trunk/core/src/main/java/org/apache/mahout/classifier/sgd/AdaptiveLogisticRegression.java (original)
+++ mahout/trunk/core/src/main/java/org/apache/mahout/classifier/sgd/AdaptiveLogisticRegression.java Mon Sep 20 22:26:07 2010
@@ -58,6 +58,8 @@ import java.util.concurrent.ExecutionExc
  * a different fitness value in non-binary cases.
  */
 public class AdaptiveLogisticRegression implements OnlineLearner {
+  private static final int SURVIVORS = 2;
+
   private int record = 0;
   private int evaluationInterval = 1000;
 
@@ -122,7 +124,15 @@ public class AdaptiveLogisticRegression 
       throw new IllegalStateException(e);
     }
 
-    ep.mutatePopulation(2);
+    // evolve based on new fitness
+    ep.mutatePopulation(SURVIVORS);
+
+    // now grossly hack the top survivors so they stick around.  Set their
+    // mutation rates small and also hack their learning rate to be small
+    // as well.
+    for (State<Wrapper> state : ep.getPopulation().subList(0, SURVIVORS)) {
+      state.getPayload().freeze(state);
+    }
     buffer.clear();
   }
 
@@ -293,6 +303,18 @@ public class AdaptiveLogisticRegression 
       wrapped.decayExponent(0);
     }
 
+    public void freeze(State<Wrapper> s) {
+      // radically decrease learning rate
+      s.getParams()[1] -= 5;
+
+      // and cause evolution to hold (almost)
+      s.setOmni(s.getOmni() / 10);
+      double[] step = s.getStep();
+      for (int i = 0; i < step.length; i++) {
+        step[i] /= 10;
+      }
+    }
+
     public void setMappings(State<Wrapper> x) {
       int i = 0;
       // set the range for regularization (lambda)