Update to gymnasium (#82)

* update install_requires

* §

* fix imports: gym → gymnasium

* fix step function

* fix step

* gym->gymnasium

* §

* fix

* fix

* fix

* fix

* fix

* fix render

* .

* fix all step returns

* Delete build/lib/carl directory

* Delete myenv directory

* Update README.md

Fix documentation link.

* make pre-commit and format

* rename every occurrence of trunched to truncated

* @Arman717

* change trunched to truncated

* Fix gymnasium version with box2d

---------

Co-authored-by: C. Benjamins <75323339+benjamc@users.noreply.github.com>

carl/envs/rna/carl_rna_definitions.py

@@ -1,5 +1,5 @@
 import numpy as np
-from gym import spaces
+from gymnasium import spaces
 
 DEFAULT_CONTEXT = {
     "mutation_threshold": 5,

carl/envs/rna/rna_environment.py

@@ -12,7 +12,7 @@
 import numpy as np
 
 from RNA import fold
-import gym
+import gymnasium as gym
 
 from typing import Any, List
 

setup.py

@@ -72,9 +72,7 @@ def read_file(filepath: str) -> str:
     include_package_data=True,
     python_requires=">=3.9",
     install_requires=[
-        "gym",
         "gymnasium>=0.27.1",
-        "pygame==2.1.0",
         "scipy>=1.7.0",
         "ConfigArgParse>=1.5.1",
         "numpy>=1.19.5",

test/test_CARLEnv.py

@@ -8,14 +8,148 @@
 
 
 class TestStateConstruction(unittest.TestCase):
-    def test_observation(self):
-        env = CARLPendulum()
-        context = CARLPendulum.get_default_context()
-        obs, info = env.reset()
+    def test_hiddenstate(self):
+        """
+        Test if we set hide_context = True that we get the original, normal state.
+        """
+        env = CARLPendulumEnv(
+            contexts={},
+            hide_context=True,
+            add_gaussian_noise_to_context=False,
+            gaussian_noise_std_percentage=0.01,
+            state_context_features=None,
+        )
+        env.reset()
+        action = [0.01]  # torque
+        state, reward, terminated, truncated, info = env.step(action=action)
+        env.close()
+        self.assertEqual(3, len(state))
+
+    def test_visiblestate(self):
+        """
+        Test if we set hide_context = False and state_context_features=None that we get the normal state extended by
+        all context features.
+        """
+        env = CARLPendulumEnv(
+            contexts={},
+            hide_context=False,
+            add_gaussian_noise_to_context=False,
+            gaussian_noise_std_percentage=0.01,
+            state_context_features=None,
+        )
+        env.reset()
+        action = [0.01]  # torque
+        state, reward, terminated, truncated, info = env.step(action=action)
+        env.close()
+        self.assertEqual(10, len(state))
+
+    def test_visiblestate_customnone(self):
+        """
+        Test if we set hide_context = False and state_context_features="changing_context_features" that we get the
+        normal state, not extended by context features.
+        """
+        env = CARLPendulumEnv(
+            contexts={},
+            hide_context=False,
+            add_gaussian_noise_to_context=False,
+            gaussian_noise_std_percentage=0.01,
+            state_context_features=["changing_context_features"],
+        )
+        env.reset()
+        action = [0.01]  # torque
+        state, reward, terminated, truncated, info = env.step(action=action)
+        env.close()
+        # Because we don't change any context features the state length should be 3
+        self.assertEqual(3, len(state))
+
+    def test_visiblestate_custom(self):
+        """
+        Test if we set hide_context = False and state_context_features=["g", "m"] that we get the
+        normal state, extended by the context feature values of g and m.
+        """
+        env = CARLPendulumEnv(
+            contexts={},
+            hide_context=False,
+            add_gaussian_noise_to_context=False,
+            gaussian_noise_std_percentage=0.01,
+            state_context_features=["g", "m"],
+        )
+        env.reset()
+        action = [0.01]  # torque
+        state, reward, terminated, truncated, info = env.step(action=action)
+        env.close()
+        # state should be of length 5 because we add two context features
+        self.assertEqual(5, len(state))
+
+    def test_visiblestate_changingcontextfeatures_nochange(self):
+        """
+        Test if we set hide_context = False and state_context_features="changing_context_features" that we get the
+        normal state, extended by the context features which are changing in the set of contexts. Here: None are
+        changing.
+        """
+        contexts = {
+            "0": {"max_speed": 8.0, "dt": 0.05, "g": 10.0, "m": 1.0, "l": 1.0},
+            "1": {"max_speed": 8.0, "dt": 0.05, "g": 10.0, "m": 1.0, "l": 1.0},
+            "2": {"max_speed": 8.0, "dt": 0.05, "g": 10.0, "m": 1.0, "l": 1.0},
+            "3": {"max_speed": 8.0, "dt": 0.05, "g": 10.0, "m": 1.0, "l": 1.0},
+        }
+        env = CARLPendulumEnv(
+            contexts=contexts,
+            hide_context=False,
+            add_gaussian_noise_to_context=False,
+            gaussian_noise_std_percentage=0.01,
+            state_context_features=["changing_context_features"],
+        )
+        env.reset()
+        action = [0.01]  # torque
+        state, reward, terminated, truncated, info = env.step(action=action)
+        env.close()
+        # state should be of length 3 because all contexts are the same
+        self.assertEqual(3, len(state))
+
+    def test_visiblestate_changingcontextfeatures_change(self):
+        """
+        Test if we set hide_context = False and state_context_features="changing_context_features" that we get the
+        normal state, extended by the context features which are changing in the set of contexts.
+        Here: Two are changing.
+        """
+        contexts = {
+            "0": {"max_speed": 8.0, "dt": 0.03, "g": 10.0, "m": 1.0, "l": 1.0},
+            "1": {"max_speed": 8.0, "dt": 0.05, "g": 10.0, "m": 1.0, "l": 0.95},
+            "2": {"max_speed": 8.0, "dt": 0.05, "g": 10.0, "m": 1.0, "l": 0.3},
+            "3": {"max_speed": 8.0, "dt": 0.05, "g": 10.0, "m": 1.0, "l": 1.3},
+        }
+        env = CARLPendulumEnv(
+            contexts=contexts,
+            hide_context=False,
+            add_gaussian_noise_to_context=False,
+            gaussian_noise_std_percentage=0.01,
+            state_context_features=["changing_context_features"],
+        )
+        env.reset()
+        action = [0.01]  # torque
+        state, reward, terminated, truncated, info = env.step(action=action)
+        env.close()
+        # state should be of length 5 because two features are changing (dt and l)
+        self.assertEqual(5, len(state))
+
+    def test_dict_observation_space(self):
+        contexts = {"0": {"max_speed": 8.0, "dt": 0.03, "g": 10.0, "m": 1.0, "l": 1.0}}
+        env = CARLPendulumEnv(
+            contexts=contexts,
+            hide_context=False,
+            dict_observation_space=True,
+            add_gaussian_noise_to_context=False,
+            gaussian_noise_std_percentage=0.01,
+            state_context_features=["changing_context_features"],
+        )
+        obs = env.reset()
         self.assertEqual(type(obs), dict)
         self.assertTrue("state" in obs)
         self.assertTrue("context" in obs)
-        self.assertEqual(len(obs["context"]), len(context))
+        action = [0.01]  # torque
+        next_obs, reward, terminated, truncated, info = env.step(action=action)
+        env.close()
 
     def test_observation_emptycontext(self):
         env = CARLPendulum(obs_context_features=[])
@@ -29,10 +163,34 @@ def test_observation_reducedcontext(self):
         state, info = env.reset()
         self.assertEqual(len(state["context"]), n)
 
-    def test_get_context_key(self):
-        contexts = self.generate_contexts()
-        env = CARLPendulum(contexts=contexts)
-        self.assertEqual(env.context_id, None)
+class TestEpisodeTermination(unittest.TestCase):
+    def test_episode_termination(self):
+        """
+        Test if we set hide_context = True that we get the original, normal state.
+        """
+        ep_length = 100
+        env = CARLPendulumEnv(
+            contexts={},
+            hide_context=True,
+            add_gaussian_noise_to_context=False,
+            gaussian_noise_std_percentage=0.01,
+            state_context_features=None,
+            max_episode_length=ep_length,
+        )
+        env.reset()
+        action = [0.0]  # torque
+        done = False
+        counter = 0
+        while not done:
+            state, reward, terminated, truncated, info = env.step(action=action)
+            counter += 1
+            self.assertTrue(counter <= ep_length)
+            if terminated or truncated:
+                done = True
+
+            if counter > ep_length:
+                break
+        env.close()
 
 
 class TestContextSampler(unittest.TestCase):
@@ -56,7 +214,7 @@ def test_context_feature_scaling_by_mean(self):
         )
         env.reset()
         action = [0.0]
-        state, reward, done, info = env.step(action=action)
+        state, reward, terminated, truncated, info = env.step(action=action)
         n_c = len(env.default_context)
         scaled_contexts = state[-n_c:]
         target = np.array(
@@ -77,7 +235,26 @@ def test_context_feature_scaling_by_default(self):
             "initial_angle_max": np.pi,
             "initial_velocity_max": 1,
         }
-        self.assertDictEqual(defaults, DEFAULT_CONTEXT)
+        contexts = {
+            "0": {"max_speed": 8.0, "dt": 0.03, "g": 10.0, "m": 1.0, "l": 1.8},
+        }
+        env = CARLPendulumEnv(
+            contexts=contexts,
+            hide_context=False,
+            add_gaussian_noise_to_context=False,
+            gaussian_noise_std_percentage=0.01,
+            state_context_features=None,
+            scale_context_features="by_default",
+            default_context=default_context,
+        )
+        env.reset()
+        action = [0.0]
+        state, reward, terminated, truncated, info = env.step(action=action)
+        n_c = len(default_context)
+        scaled_contexts = state[-n_c:]
+        self.assertTrue(
+            np.all(np.array([1.0, 0.6, 1, 1, 1.8, 1, 1]) == scaled_contexts)
+        )
 
     def test_sample_contexts(self):
         from carl.context.sampling import sample_contexts