improved more compact cbify

1 files changed, 41 insertions, 12 deletions

diff --git a/vowpalwabbit/cbify.cc b/vowpalwabbit/cbify.cc
index 2bfcbef5..8061f383 100644
--- a/vowpalwabbit/cbify.cc
+++ b/vowpalwabbit/cbify.cc
@@ -165,6 +165,26 @@ namespace CBIFY {
     return 1;
   }
   
+  void safety(v_array<float>& distribution, float min_prob)
+  {
+    float added_mass = 0.;
+    for (uint32_t i = 0; i < distribution.size();i++)
+      if (distribution[i] > 0 && distribution[i] <= min_prob)
+	{
+	  added_mass += min_prob - distribution[i];
+	  distribution[i] = min_prob;
+	}
+    
+    float ratio = 1. / (1. + added_mass);
+    if (ratio < 0.999)
+      {
+	for (uint32_t i = 0; i < distribution.size(); i++)
+	  if (distribution[i] > min_prob)
+	    distribution[i] = distribution[i] * ratio; 
+	safety(distribution, min_prob);
+      }
+  }
+
   void gen_cs_label(vw& all, CB::cb_class& known_cost, example& ec, CSOAA::label& cs_ld, uint32_t label)
   {
     CSOAA::wclass wc;
@@ -188,14 +208,12 @@ namespace CBIFY {
     //Use cost sensitive oracle to cover actions to form distribution.
     OAA::mc_label* ld = (OAA::mc_label*)ec.ld;
     data.counter++;
-    float round_epsilon = data.epsilon / sqrt(data.counter);
 
-    float base_prob = round_epsilon / data.k;
     data.count.erase();
     data.cs_label.costs.erase();
     for (size_t j = 0; j < data.k; j++)
       {
-	data.count.push_back(base_prob);
+	data.count.push_back(0);
 
 	CSOAA::wclass wc;
 	
@@ -207,18 +225,26 @@ namespace CBIFY {
 	data.cs_label.costs.push_back(wc);
       }
 
-    float additive_probability = (1. - round_epsilon) / (float)data.bags;
+    float additive_probability = 1. / (float)data.bags;
 
     ec.ld = &data.cs_label;
     for (size_t i = 0; i < data.bags; i++)
       { //get predicted cost-sensitive predictions
-	data.cs->predict(ec,i+2);
+	if (i == 0)
+	  data.cs->predict(ec, i);
+	else
+	  data.cs->predict(ec,i+1);
 	data.count[ec.final_prediction-1] += additive_probability;
 	data.predictions[i] = ec.final_prediction;
       }
+
+    float min_prob = data.epsilon * min (1. / data.k, 1. / sqrt(data.counter * data.k));
+    
+    safety(data.count, min_prob);
+    
     //compute random action
     uint32_t action = choose_action(data.count);
-
+    
     if (is_learn)
       {
 	data.cb_label.costs.erase();
@@ -233,12 +259,11 @@ namespace CBIFY {
 	
 	//1. Compute loss vector
 	data.cs_label.costs.erase();
-	float norm = 0;
+	float norm = min_prob * data.k;
 	for (uint32_t j = 0; j < data.k; j++)
 	  { //data.cs_label now contains an unbiased estimate of cost of each class.
 	    gen_cs_label(*data.all, l, ec, data.cs_label, j+1);
-	    data.count[j] = base_prob;
-	    norm += base_prob;
+	    data.count[j] = 0;
 	  }
 	
 	ec.ld = &data.second_cs_label;
@@ -247,13 +272,17 @@ namespace CBIFY {
 	  { //get predicted cost-sensitive predictions
 	    for (size_t j = 0; j < data.k; j++)
 	      {
-		float pseudo_cost = data.cs_label.costs[j].x - 0.125 * (base_prob / (data.count[j] / norm) + 1);
+		float pseudo_cost = data.cs_label.costs[j].x - data.epsilon * min_prob / (max(data.count[j], min_prob) / norm) + 1;
 		data.second_cs_label.costs[j].weight_index = j+1;
 		data.second_cs_label.costs[j].x = pseudo_cost;
 	      }
-	    data.cs->learn(ec,i+2);
+	    if (i != 0)
+	      data.cs->learn(ec,i+1);
+	    if (data.count[data.predictions[i]-1] < min_prob)
+	      norm += max(0, additive_probability - (min_prob - data.count[data.predictions[i]-1]));
+	    else
+	      norm += additive_probability;
 	    data.count[data.predictions[i]-1] += additive_probability;
-	    norm += additive_probability;
 	  }
       }