[RLlib] Fix DDPG learning/release tests and make MARWIL CI test crite…

…rium more difficult. (#39386)
ray-project · Sep 7, 2023 · c324f38 · c324f38
1 parent f33b8eb
commit c324f38
Show file tree

Hide file tree

Showing 2 changed files with 9 additions and 8 deletions.
diff --git a/...tests/yaml_files/ddpg/ddpg-hopper-v4.yaml → .../yaml_files/ddpg/ddpg-halfcheetah-v4.yaml b/...tests/yaml_files/ddpg/ddpg-hopper-v4.yaml → .../yaml_files/ddpg/ddpg-halfcheetah-v4.yaml
@@ -1,12 +1,12 @@
-ddpg-hopper-v4:
-    env: Hopper-v4
+ddpg-halfcheetah-v4:
+    env: HalfCheetah-v4
     run: DDPG
     # Minimum reward and total ts (in given time_total_s) to pass this test.
     pass_criteria:
-        sampler_results/episode_reward_mean: 110.0
-        timesteps_total: 50000
+        sampler_results/episode_reward_mean: 1000.0
+        timesteps_total: 100000
     stop:
-        time_total_s: 1800
+        time_total_s: 3600
     config:
         actor_hiddens: [256, 256]
         critic_hiddens: [256, 256]

diff --git a/rllib/algorithms/marwil/tests/test_marwil.py b/rllib/algorithms/marwil/tests/test_marwil.py
@@ -55,15 +55,16 @@ def test_marwil_compilation_and_learning_from_offline_file(self):
                 evaluation_parallel_to_training=True,
                 evaluation_config=marwil.MARWILConfig.overrides(input_="sampler"),
                 off_policy_estimation_methods={},
+                always_attach_evaluation_results=True,
             )
             .offline_data(input_=[data_file])
         )
 
         num_iterations = 350
-        min_reward = 70.0
+        min_reward = 100.0
 
         # Test for all frameworks.
-        for _ in framework_iterator(config, frameworks=("tf", "torch")):
+        for _ in framework_iterator(config, frameworks=("torch", "tf")):
             algo = config.build()
             learnt = False
             for i in range(num_iterations):
@@ -127,7 +128,7 @@ def test_marwil_cont_actions_from_offline_file(self):
         num_iterations = 3
 
         # Test for all frameworks.
-        for _ in framework_iterator(config, frameworks=("tf", "torch")):
+        for _ in framework_iterator(config, frameworks=("torch", "tf")):
             algo = config.build(env="Pendulum-v1")
             for i in range(num_iterations):
                 print(algo.train())