Fix the bug of having -100000 or 10000 rewards sometime.

backends/controller_base.py

@@ -65,7 +65,7 @@ class ControllerBase(PolicyBase):
         total_reward += self.current_node.high_level_extra_reward
 
         # TODO for info
-        return observation, total_reward, self.env.termination_condition, info
+        return observation, total_reward, terminal, info
 
     # TODO: Looks generic. Move to an intermediate class/highlevel manager so that base class can be clean
     ''' Executes one step of current node. Sets node_terminal_state_reached flag if node termination condition
@@ -76,9 +76,7 @@ class ControllerBase(PolicyBase):
 
     def low_level_step_current_node(self):
 
-        u_ego = self.current_node.low_level_policy(
-            self.current_node.get_reduced_features_tuple())
+        u_ego = self.current_node.low_level_policy(self.current_node.get_reduced_features_tuple())
         feature, R, terminal, info = self.current_node.step(u_ego)
         self.node_terminal_state_reached = terminal
-        return self.env.get_features_tuple(
-        ), R, self.env.termination_condition, info
+        return self.env.get_features_tuple(), R, self.env.termination_condition, info

backends/kerasrl_learner.py

@@ -317,6 +317,7 @@ class DQNLearner(LearnerBase):
             nb_steps_warmup=self.nb_steps_warmup,
             target_model_update=self.target_model_update,
             policy=policy,
+            test_policy=test_policy,
             enable_dueling_network=True)
 
         agent.compile(Adam(lr=self.lr), metrics=['mae'])
@@ -327,6 +328,8 @@ class DQNLearner(LearnerBase):
               env,
               nb_steps=1000000,
               visualize=False,
+              verbose=1,
+              log_interval=10000,
               nb_max_episode_steps=200,
               tensorboard=False,
               model_checkpoints=False,
@@ -346,7 +349,8 @@ class DQNLearner(LearnerBase):
             env,
             nb_steps=nb_steps,
             visualize=visualize,
-            verbose=1,
+            verbose=verbose,
+            log_interval=log_interval,
             nb_max_episode_steps=nb_max_episode_steps,
             callbacks=callbacks)
 
@@ -479,21 +483,13 @@ class RestrictedGreedyQPolicy(GreedyQPolicy):
             Selection action
         """
         assert q_values.ndim == 1
-        nb_actions = q_values.shape[0]
-        restricted_q_values = list()
-
-        for i in range(0, nb_actions):
-            if q_values[i] != -np.inf:
-                restricted_q_values.append(q_values[i])
 
-        # every q_value is -np.inf (this sometimes inevitably happens within the fit and test functions
-        # of kerasrl at the terminal stage as they force to call forward in Kerasrl-learner which calls this function.
-        # In this case, we choose a policy randomly.
-        if len(restricted_q_values) < 1:
-            action = np.random.random_integers(0, nb_actions - 1)
+        if np.max(q_values) == - np.inf:
+            # every q_value is -np.inf, we choose action = 0
+            action = 0
 
         else:
-            action = np.argmax(restricted_q_values)
+            action = np.argmax(q_values)
 
         return action
 
@@ -521,6 +517,7 @@ class DQNAgentOverOptions(DQNAgent):
     def __get_invalid_node_indices(self):
         """Returns a list of option indices that are invalid according to
         initiation conditions."""
+
         invalid_node_indices = list()
         for index, option_alias in enumerate(self.low_level_policy_aliases):
             # TODO: Locate reset_maneuver to another place as this is a "get" function.

config.json

@@ -3,7 +3,7 @@
 		"wait": "Wait",
 		"follow": "Follow",
 		"stop": "Stop",
-		"immediatestop": "ImmediateStop",
+		"changelane": "ChangeLane",
 		"keeplane": "KeepLane"
 	},
 

env/env_base.py

@@ -157,7 +157,7 @@ class EpisodicEnvBase(GymCompliantEnvBase):
         if self._terminate_in_goal and self.goal_achieved:
             return True
 
-        return self.violation_happened and self._LTL_preconditions_enable
+        return self.violation_happened
 
     @property
     def goal_achieved(self):

env/simple_intersection/road_geokinemetry.py

@@ -93,7 +93,7 @@ vlanes = Route(
     [-vwidth - 5.0 - intersection_voffset, -vwidth - intersection_voffset], 35,
     [-4.0, 4.0])
 
-intersection_width = vlanes.n_lanes * vlanes.width
+intersection_width =  vlanes.n_lanes * vlanes.width
 intersection_height = hlanes.n_lanes * hlanes.width
 
 intersection_width_w_offset = intersection_width + 2 * intersection_hoffset

env/simple_intersection/simple_intersection_env.py

@@ -51,7 +51,9 @@ class SimpleIntersectionEnv(RoadEnv, EpisodicEnvBase):
 
     #: The weight vector to calculate the cost. In the maneuver, cost_weights
     # can be set to a specific value which may be different than the default.
-    cost_weights = (1.0, 0.25, 0.1, 1.0, 100.0, 0.1, 0.25, 0.1)
+    # TODO: check _cost_weights in both here and ManeuverBase. The _cost_weights has to be substituted to here, but it doesn't sometimes.
+    _cost_weights = (10.0 * 1e-3, 10.0 * 1e-3, 0.25 * 1e-3, 1.0 * 1e-3,
+                     100.0 * 1e-3, 0.1 * 1e-3, 0.05 * 1e-3, 0.1 * 1e-3)
 
     #TODO: Move this to constants
     # The empirical min and max of each term in the cost vector, which is used to normalize the values

options/options_loader.py

@@ -180,11 +180,6 @@ class OptionsGraph:
                 print("\n Warning: the trained low-level policy of \"" + key +
                       "\" does not exists; the manual policy will be used.\n")
 
-            # setting the cost weights for high-level policy training.
-            # TODO: this shouldn't be initialized here, but within ManeuverBase class (e.g. make some flag indicating high-level training and set the weights)...
-            maneuver._cost_weights = (100.0 * 1e-3, 10.0 * 1e-3, 0.25 * 1e-3, 1.0 * 1e-3,
-                                     100.0 * 1e-3, 0.1 * 1e-3, 0.25 * 1e-3, 0.1 * 1e-3)
-
             if self.config["method"] == "mcts":
                 maneuver.timeout = np.inf
 

options/simple_intersection/maneuver_base.py

@@ -30,7 +30,7 @@ class ManeuverBase(EpisodicEnvBase):
     # as a negative reward, so a cost will be summed up to the reward
     # with subtraction.
     _cost_weights = (10.0 * 1e-3, 10.0 * 1e-3, 0.25 * 1e-3, 1.0 * 1e-3,
-                     100.0 * 1e-3, 0.1 * 1e-3, 0.25 * 1e-3, 0.1 * 1e-3)
+                     100.0 * 1e-3, 0.1 * 1e-3, 0.05 * 1e-3, 0.1 * 1e-3)
 
     _extra_r_terminal = None
     _extra_r_on_timeout = None
@@ -38,7 +38,7 @@ class ManeuverBase(EpisodicEnvBase):
     #: the flag being False when _cost_weights is used without
     # modification; If True, then the action parts of _cost_weights
     # are increased for some edge cases (see the step method).
-    _extra_action_weights_flag = True
+    _extra_action_weights_flag = False
 
     #: the extra weights on the actions added to _cost_weights
     # for some edge cases when _extra_action_weights_flag = True.