Return deterministic actions for training (#5615)

* Added more stable test.

* Fix the tests.

* Fix pre-commit

* Fix help line to pass precommit.

Author: cmard

ml-agents/mlagents/trainers/cli_utils.py

@@ -96,7 +96,8 @@ def _create_parser() -> argparse.ArgumentParser:
         default=False,
         dest="deterministic",
         action=DetectDefaultStoreTrue,
-        help="Whether to select actions deterministically in policy. `dist.mean` for continuous action space, and `dist.argmax` for deterministic action space ",
+        help="Whether to select actions deterministically in policy. `dist.mean` for continuous action "
+        "space, and `dist.argmax` for deterministic action space ",
     )
     argparser.add_argument(
         "--force",

ml-agents/mlagents/trainers/tests/torch/test_action_model.py

@@ -65,24 +65,22 @@ def test_deterministic_sample_action():
     agent_action2 = action_model._sample_action(dists)
     agent_action3 = action_model._sample_action(dists)
 
-    chance_counter = 0
-
-    if not torch.equal(
+    assert not torch.equal(
         agent_action1.continuous_tensor, agent_action2.continuous_tensor
-    ):
-        chance_counter += 1
+    )
 
-    if not torch.equal(
+    assert not torch.equal(
         agent_action1.continuous_tensor, agent_action3.continuous_tensor
-    ):
-        chance_counter += 1
+    )
 
-    assert chance_counter > 1
     chance_counter = 0
     if not torch.equal(agent_action1.discrete_tensor, agent_action2.discrete_tensor):
         chance_counter += 1
     if not torch.equal(agent_action1.discrete_tensor, agent_action3.discrete_tensor):
         chance_counter += 1
+    if not torch.equal(agent_action2.discrete_tensor, agent_action3.discrete_tensor):
+        chance_counter += 1
+
     assert chance_counter > 1