add epsilon-greedy for dqn

examples/dqn_example.py

@@ -66,9 +66,11 @@ if __name__ == '__main__':
         pi.sync_weights()  # TODO: automate this for policies with target network
 
         start_time = time.time()
-        for i in range(100):
+        #TODO : repeat_num shoulde be defined in some configuration files
+        repeat_num = 100
+        for i in range(repeat_num):
             # collect data
-            data_collector.collect(num_episodes=50)
+            data_collector.collect(num_episodes=50, epsilon_greedy= (repeat_num - i + 0.0) / repeat_num)
 
             # print current return
             print('Epoch {}:'.format(i))

tianshou/core/policy/dqn.py

@@ -18,7 +18,7 @@ class DQN(PolicyBase):
         else:
             self.interaction_count = -1
 
-    def act(self, observation, exploration=None):
+    def act(self, observation, my_feed_dict):
         sess = tf.get_default_session()
         if self.weight_update > 1:
             if self.interaction_count % self.weight_update == 0:
@@ -30,7 +30,6 @@ class DQN(PolicyBase):
         if self.weight_update > 0:
             self.interaction_count += 1
 
-        if not exploration:
         return np.squeeze(action)
 
     @property

tianshou/data/batch.py

@@ -34,7 +34,7 @@ class Batch(object):
         self._is_first_collect = True
 
     def collect(self, num_timesteps=0, num_episodes=0, my_feed_dict={},
-                process_reward=True):  # specify how many data to collect here, or fix it in __init__()
+                process_reward=True, epsilon_greedy=0):  # specify how many data to collect here, or fix it in __init__()
         assert sum(
             [num_timesteps > 0, num_episodes > 0]) == 1, "One and only one collection number specification permitted!"
 
@@ -106,6 +106,10 @@ class Batch(object):
                 episode_start_flags.append(True)
 
                 while True:
+                    # a simple implementation of epsilon greedy
+                    if epsilon_greedy > 0 and np.random.random() < epsilon_greedy:
+                        ac = np.random.randint(low = 0, high = self._env.action_space.n)
+                    else:
                         ac = self._pi.act(ob, my_feed_dict)
                     actions.append(ac)
 
@@ -114,9 +118,9 @@ class Batch(object):
                     ob, reward, done, _ = self._env.step(ac)
                     rewards.append(reward)
 
-                    t_count += 1
+                    #t_count += 1
-                    if t_count >= 100:  # force episode stop, just to test if memory still grows
+                    #if t_count >= 100:  # force episode stop, just to test if memory still grows
-                        break
+                    #    break
 
                     if done:  # end of episode, discard s_T
                         # TODO: for num_timesteps collection, has to store terminal flag instead of start flag!