Support for custom observations (both observation function and observ…

…ation space) (#133)

* Minor style changes to my previous implementation (to match the author's style)

* Support for custom observation function, preparation for drq_norm

* Working support for registerable observation functions, minor implementation changes

* Moved checking the reward function to TrafficSignal constructor

* Observations moved to another file

* Minor fixes regarding observations implementation

sumo_rl/__init__.py

@@ -1,3 +1,3 @@
-from sumo_rl.environment.env import SumoEnvironment, TrafficSignal
+from sumo_rl.environment.env import SumoEnvironment, TrafficSignal, ObservationFunction
 from sumo_rl.environment.env import env, parallel_env
 from sumo_rl.environment.resco_envs import grid4x4, arterial4x4, ingolstadt1, ingolstadt7, ingolstadt21, cologne1, cologne3, cologne8

sumo_rl/environment/env.py

@@ -19,6 +19,7 @@
 from pettingzoo.utils.conversions import parallel_wrapper_fn
 
 from .traffic_signal import TrafficSignal
+from .observations import ObservationFunction, DefaultObservationFunction
 
 LIBSUMO = 'LIBSUMO_AS_TRACI' in os.environ
 
@@ -50,7 +51,7 @@ class SumoEnvironment(gym.Env):
     :param max_green: (int) Max green time in a phase
     :single_agent: (bool) If true, it behaves like a regular gym.Env. Else, it behaves like a MultiagentEnv (https://github.com/ray-project/ray/blob/master/python/ray/rllib/env/multi_agent_env.py)
     :reward_fn: (str/function/dict) String with the name of the reward function used by the agents, a reward function, or dictionary with reward functions assigned to individual traffic lights by their keys
-    :observation_fn: (str/function) String with the name of the observation function or a callable observation function itself
+    :observation_class: (ObservationFunction) Inherited class which has both the observation function and observation space
     :add_system_info: (bool) If true, it computes system metrics (total queue, total waiting time, average speed) in the info dictionary
     :add_per_agent_info: (bool) If true, it computes per-agent (per-traffic signal) metrics (average accumulated waiting time, average queue) in the info dictionary
     :sumo_seed: (int/string) Random seed for sumo. If 'random' it uses a randomly chosen seed.
@@ -83,7 +84,7 @@ def __init__(
         max_green: int = 50, 
         single_agent: bool = False,
         reward_fn: Union[str,Callable,dict] = 'diff-waiting-time',
-        observation_fn: Union[str,Callable] = 'default',
+        observation_class: ObservationFunction = DefaultObservationFunction,
         add_system_info: bool = True,
         add_per_agent_info: bool = True,
         sumo_seed: Union[str,int] = 'random', 
@@ -135,7 +136,7 @@ def __init__(
             traci.start([sumolib.checkBinary('sumo'), '-n', self._net], label='init_connection'+self.label)
             conn = traci.getConnection('init_connection'+self.label)
         self.ts_ids = list(conn.trafficlight.getIDList())
-        self.observation_fn = observation_fn
+        self.observation_class = observation_class
 
         if isinstance(self.reward_fn, dict):
             self.traffic_signals = {ts: TrafficSignal(self,

sumo_rl/environment/observations.py

@@ -0,0 +1,45 @@
+from .traffic_signal import TrafficSignal
+from abc import abstractmethod
+from gymnasium import spaces
+import numpy as np
+
+class ObservationFunction:
+    """
+    Abstract base class for observation functions.
+    """
+    def __init__(self, ts: TrafficSignal):
+        self.ts = ts
+
+    @abstractmethod
+    def __call__(self):
+        """
+        Subclasses must override this method.
+        """
+        pass
+
+    @abstractmethod
+    def observation_space(self):
+        """
+        Subclasses must override this method.
+        """
+        pass
+
+
+class DefaultObservationFunction(ObservationFunction):
+    def __init__(self, ts: TrafficSignal):
+        super().__init__(ts)
+
+    def __call__(self):
+        phase_id = [1 if self.ts.green_phase == i else 0 for i in range(self.ts.num_green_phases)]  # one-hot encoding
+        min_green = [0 if self.ts.time_since_last_phase_change < self.ts.min_green + self.ts.yellow_time else 1]
+        density = self.ts.get_lanes_density()
+        queue = self.ts.get_lanes_queue()
+        observation = np.array(phase_id + min_green + density + queue, dtype=np.float32)
+        return observation
+
+    def observation_space(self):
+        return spaces.Box(
+            low=np.zeros(self.ts.num_green_phases+1+2*len(self.ts.lanes), dtype=np.float32), 
+            high=np.ones(self.ts.num_green_phases+1+2*len(self.ts.lanes), dtype=np.float32)
+        )
+

sumo_rl/environment/traffic_signal.py

@@ -60,13 +60,7 @@ def __init__(self,
             else:
                 raise NotImplementedError(f'Reward function {self.reward_fn} not implemented')
 
-        if isinstance(self.env.observation_fn, Callable):
-            self.observation_fn = self.env.observation_fn
-        else:
-            if self.env.observation_fn in TrafficSignal.observation_fns.keys():
-                self.observation_fn = TrafficSignal.observation_fns[self.env.observation_fn]
-            else:
-                raise NotImplementedError(f'Observation function {self.env.observation_fn} not implemented')
+        self.observation_fn = self.env.observation_class(self)
 
         self.build_phases()
 
@@ -75,7 +69,7 @@ def __init__(self,
         self.out_lanes = list(set(self.out_lanes))
         self.lanes_lenght = {lane: self.sumo.lane.getLength(lane) for lane in self.lanes + self.out_lanes}
 
-        self.observation_space = spaces.Box(low=np.zeros(self.num_green_phases+1+2*len(self.lanes), dtype=np.float32), high=np.ones(self.num_green_phases+1+2*len(self.lanes), dtype=np.float32))
+        self.observation_space = self.observation_fn.observation_space()
         self.discrete_observation_space = spaces.Tuple((
             spaces.Discrete(self.num_green_phases),                       # Green Phase
             spaces.Discrete(2),                                           # Binary variable active if min_green seconds already elapsed
@@ -148,7 +142,7 @@ def set_next_phase(self, new_phase):
             self.time_since_last_phase_change = 0
 
     def compute_observation(self):
-        return self.observation_fn(self)
+        return self.observation_fn()
 
     def compute_reward(self):
         self.last_reward = self.reward_fn(self)
@@ -233,20 +227,9 @@ def register_reward_fn(cls, fn):
 
         cls.reward_fns[fn.__name__] = fn
 
-    @classmethod
-    def register_observation_fn(cls, fn):
-        if fn.__name__ in cls.observation_fns.keys():
-            raise KeyError(f'Observation function {fn.__name__} already exists')
-
-        cls.observation_fns[fn.__name__] = fn
-
     reward_fns = {
         'diff-waiting-time': _diff_waiting_time_reward,
         'average-speed': _average_speed_reward,
         'queue': _queue_reward,
         'pressure': _pressure_reward
-    }
-
-    observation_fns = {
-        'default': _observation_fn_default
     }