Revert some changes in simple_intersection_env and maneuver_base.

env/simple_intersection/simple_intersection_env.py

@@ -256,7 +256,7 @@ class SimpleIntersectionEnv(RoadEnv, EpisodicEnvBase):
         ego_heading_towards_lane_centre = \
             kwargs.setdefault('ego_heading_towards_lane_centre', True)
         randomize_special_scenarios = \
-            kwargs.setdefault('randomize_special_scenarios', True)
+            kwargs.setdefault('randomize_special_scenarios', False)
 
         if n_others_range[0] < 0:
             raise ValueError("the number of other vehicles has to be non-negative.")

options/simple_intersection/maneuver_base.py

@@ -29,7 +29,7 @@ class ManeuverBase(EpisodicEnvBase):
     # _extra_action_weights_flag = True); note that a cost is defined
     # as a negative reward, so a cost will be summed up to the reward
     # with subtraction.
-    _cost_weights = ( 20.0*1e-3,  1.0*1e-3, 0.25*1e-3, 1.0*1e-3,
+    _cost_weights = (  1.0*1e-3,  1.0*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)
 
     _extra_r_terminal = None
@@ -49,7 +49,7 @@ class ManeuverBase(EpisodicEnvBase):
 
     #: enable the property for low-level policy training only, if any.
     # this flag is useful in low-level policy learning
-    # (see maneuvers.py. In ManeuverBase, this plays no role).
+    # (see maneuvers.py).
     _enable_low_level_training_properties = False
 
     #: the additional reward given to the high-level learner for choosing