homework7

homework7/LinUCB-Homework.py

@@ -85,11 +85,11 @@ class UCBArm(object):
     """
     """
     def update_alpha(self, method=1, t=None):
     def update_alpha(self, method=1, t=None):
         if method == 1:
         if method == 1:
-            self.alpha = 0.1 / np.sqrt(t + 1)
+            self.alpha = 1 / np.sqrt(t)
         elif method == 2:
         elif method == 2:
             self.alpha = 0.1
             self.alpha = 0.1
         elif method == 3:
         elif method == 3:
-            self.alpha = 10 / (np.sqrt(t + 1) + 2 * t)
+            self.alpha = 10 / (np.sqrt(t) + t)
         return None
         return None
 
 
 
 
@@ -203,8 +203,6 @@ class LinUCB(object):
 
 
 
 
 # In[19]:
 # In[19]:
-
-
 """
 """
 Class LinUCB implements unbiased offline evaluation
 Class LinUCB implements unbiased offline evaluation
 of our multi-arm contextual bandit following
 of our multi-arm contextual bandit following
@@ -243,14 +241,17 @@ class bandit_evaluator(object):
     """
     """
     def calc_ctr(self, x, action, t):
     def calc_ctr(self, x, action, t):
         assert t > 0
         assert t > 0
-        bandit = self.bandits[action]
-        pred_act = bandit.predict[x]
+        bandit = self.bandits[-1]
+        pred_act = bandit.predict(x)
         ### todo
         ### todo
-        ctr =
+        if len(self.ctr_history):
+            hist = self.ctr_history[-1] * (t - 1)
+        else:
+            hist = 0
+        ctr = (hist + int(pred_act == action)) / t
         self.ctr_history.append(ctr)
         self.ctr_history.append(ctr)
         return ctr
         return ctr
 
 
-
 # In[20]:
 # In[20]:
 
 
 
 
@@ -285,8 +286,8 @@ def train(file, steps, alpha, nArms, d):
         bandit.arms[arm].update_arm(reward, x)
         bandit.arms[arm].update_arm(reward, x)
 
 
         if t > 0: # explore various alpha update methods to improve CTR
         if t > 0: # explore various alpha update methods to improve CTR
-            bandit.arms[arm].update_alpha(method=2) # or method=2
-            #bandit.arms[arm].update_alpha(3, t)
+            # bandit.arms[arm].update_alpha(method=2) # or method=2
+            bandit.arms[arm].update_alpha(3, t)
 
 
         if t > 0: # evaluate current bandit algorithm
         if t > 0: # evaluate current bandit algorithm
             evaluator.bandits.append(bandit)
             evaluator.bandits.append(bandit)