change initial weights for ucb agent

agent/__init__.py

@@ -1 +1,2 @@
-from .util import create_agent_helper
+from .util import create_agent_helper
+from .ucb_agent import UCBAgent

agent/ucb_agent.py

@@ -0,0 +1,59 @@
+from recsim.agent import AbstractEpisodicRecommenderAgent
+import tensorflow as tf
+import numpy as np
+
+class UCBAgent(AbstractEpisodicRecommenderAgent):
+  def __init__(self, sess, observation_space, action_space, eval_mode, alpha=1.0, learning_rate=0.001, summary_writer=None):
+    super(UCBAgent, self).__init__(action_space, summary_writer)
+    self._num_candidates = int(action_space.nvec[0])
+    self._W = tf.Variable(np.random.uniform(0, 10, size=(self._num_candidates, 3)), name='W')
+    self._sess = sess
+    self._return_idx = None
+    self._prev_pred_pr = None
+    self._opt = tf.compat.v1.train.GradientDescentOptimizer(learning_rate)
+    self._alpha = alpha
+  
+    assert self._slate_size == 1
+  def step(self, reward, observation):
+    docs = observation['doc']
+    user = observation['user']
+    response = observation['response']
+
+    if self._return_idx != None and response != None:
+      # update w
+      y_true = [response[0]['recall']]
+      y_pred = self._prev_pred_pr
+      loss = tf.losses.binary_crossentropy(y_true, y_pred)
+      self._sess.run(self._opt.minimize(loss))
+    base_pr = self.calc_prs(user['time'], user['last_review'], user['history'], self._W)
+
+    time = user['time'] + 1
+    history_pos = user['history'].copy()
+    history_pos[:, [0, 1]] += 1 # add n, n+ by 1
+    history_neg = user['history'].copy()
+    history_neg[:, [0, 2]] += 1 # add n, n- by 1
+    last_review_now = np.repeat(user['time'], len(user['last_review']))
+    pr_pos = self.calc_prs(time, last_review_now, history_pos, self._W)
+    pr_neg = self.calc_prs(time, last_review_now, history_neg, self._W)
+
+    gain = (pr_pos + pr_neg) / 2 - base_pr
+    time_since_last_review = user['time'] - user['last_review']
+    uncertainty = self._alpha * tf.math.sqrt(tf.math.log(time_since_last_review) / user['history'][:, 0])
+    # print(gain.eval(session=self._sess))
+    # print(time_since_last_review)
+    # print(uncertainty.eval(session=self._sess))
+    ucb_score = gain + uncertainty
+    print("       gain:", gain.eval(session=self._sess))
+    print("uncertainty:", uncertainty.eval(session=self._sess))
+    best_idx = tf.argmax(ucb_score)
+
+    self._return_idx = self._sess.run(best_idx)
+    self._prev_pred_pr = base_pr[self._return_idx]
+    return [self._return_idx]
+
+    
+  def calc_prs(self, train_time, last_review, history, W):
+    last_review = train_time - last_review
+    mem_param = tf.math.exp(tf.reduce_sum(history * W, axis=1))
+    pr = tf.math.exp(-last_review / mem_param)
+    return pr

document/FlashcardDocumentSampler.py

@@ -2,13 +2,13 @@ from .FlashcardDocument import FlashcardDocument
 from recsim import document
 
 class FlashcardDocumentSampler(document.AbstractDocumentSampler):
-  def __init__(self, doc_ctor=FlashcardDocument, seed=0, **kwargs):
-    super(FlashcardDocumentSampler, self).__init__(doc_ctor, seed, **kwargs)
+  def __init__(self, doc_ctor=FlashcardDocument, **kwargs):
+    super(FlashcardDocumentSampler, self).__init__(doc_ctor, **kwargs)
     self._doc_count = 0
 
   def sample_document(self):
     doc_features = {}
     doc_features['doc_id'] = self._doc_count
-    doc_features['difficulty'] = self._rng.uniform(0, 3, (1, 3))
+    doc_features['difficulty'] = self._rng.uniform(0, 5, (1, 3))
     self._doc_count += 1
     return self._doc_ctor(**doc_features)

main.py

@@ -17,8 +17,8 @@ tf.compat.v1.disable_eager_execution()
 create_agent_fn = create_agent_helper(full_slate_q_agent.FullSlateQAgent)
 
 ltsenv = environment.Environment(
-  FlashcardUserModel(num_candidates, time_budget, slate_size, seed=0, sample_seed=0),
-  FlashcardDocumentSampler(seed=0),
+  FlashcardUserModel(num_candidates, time_budget, slate_size),
+  FlashcardDocumentSampler(),
   num_candidates,
   slate_size,
   resample_documents=False)

user/FlashcardUserModel.py

@@ -7,14 +7,13 @@ from util import eval_result
 import numpy as np
 
 class FlashcardUserModel(user.AbstractUserModel):
-  def __init__(self, num_candidates, time_budget, slate_size, seed=0, sample_seed=0):
+  def __init__(self, num_candidates, time_budget, slate_size, seed=0):
     super(FlashcardUserModel, self).__init__(
         UserResponse, UserSampler(
             UserState, num_candidates, time_budget, 
-            seed=sample_seed
+            seed=seed
         ), slate_size)
     self.choice_model = MultinomialLogitChoiceModel({})
-    self._rng = np.random.RandomState(seed)
 
   def is_terminal(self):
     terminated = self._user_state._time > self._user_state._time_budget
@@ -53,8 +52,7 @@ class FlashcardUserModel(user.AbstractUserModel):
     doc_id = doc._doc_id
     W = self._user_state._W[doc_id]
     if not W.any(): # uninitialzed
-      error = self._user_state._doc_error[doc_id] # a uniform error for each user
-      self._user_state._W[doc_id] = W = doc.base_difficulty * error
+      self._user_state._W[doc_id] = W = doc.base_difficulty * np.random.uniform(0.5, 2.0, (1, 3)) # a uniform error for each user
       print(W)
     # use exponential function to simulate whether the user recalls
     last_review = self._user_state._time - self._user_state._last_review[doc_id]
@@ -62,6 +60,6 @@ class FlashcardUserModel(user.AbstractUserModel):
 
     pr = np.exp(-last_review / np.exp(np.dot(W, x))).squeeze()
     print(f"time: {self._user_state._time}, reviewing flashcard {doc_id}, recall rate = {pr}")
-    if self._rng.random_sample() < pr: # remembered
+    if np.random.rand() < pr: # remembered
       response._recall = True
     response._pr = pr

user/UserSampler.py

@@ -7,13 +7,9 @@ class UserSampler(user.AbstractUserSampler):
                num_candidates=10,
                time_budget=60,
                **kwargs):
+    self._state_parameters = {'num_candidates': num_candidates, 'time_budget': time_budget}
     super(UserSampler, self).__init__(user_ctor, **kwargs)
-    doc_error = self._rng.uniform(0.5, 1.5, (num_candidates, 3))
-    self._state_parameters = {
-      'num_candidates': num_candidates, 
-      'time_budget': time_budget,
-      'doc_error': doc_error
-    }
+
 
   def sample_user(self):
     return self._user_ctor(**self._state_parameters)

user/UserState.py

@@ -3,14 +3,13 @@ import numpy as np
 from gym import spaces
 
 class UserState(user.AbstractUserState):
-  def __init__(self, num_candidates, time_budget, doc_error):
+  def __init__(self, num_candidates, time_budget):
     self._cards = num_candidates
     self._history = np.zeros((num_candidates, 3))
     self._last_review = np.repeat(-1.0, num_candidates)
     self._time_budget = time_budget
     self._time = 0
     self._W = np.zeros((num_candidates, 3))
-    self._doc_error = doc_error
     super(UserState, self).__init__()
   def create_observation(self):
     return {'history': self._history, 'last_review': self._last_review, 'time': self._time, 'time_budget': self._time_budget}