Policy's action with an action mask not respecting its min/max bounds

[alfoudari]

The following unit test fails by producing actions that are out of the action_spec min/max boundaries:

https://github.com/abstractpaper/tf-agents/blob/d06b2fcf4fecb3e052224dd257ad3e1b9ef7d4b0/tf_agents/policies/q_policy_test.py#L234-L269

Running the unit test:

(.venv) aziz@merlin:~/code/agents$ python -m unittest tf_agents.policies.q_policy_test.QPolicyTest.testActionWithinBoundsWithMasking
culprits: [(2, 8), (264, 8), (822, 8), (881, 8)]
F
======================================================================
FAIL: testActionWithinBoundsWithMasking (tf_agents.policies.q_policy_test.QPolicyTest)
testActionWithinBoundsWithMasking (tf_agents.policies.q_policy_test.QPolicyTest)
----------------------------------------------------------------------
Traceback (most recent call last):
  File "/home/aziz/code/agents/tf_agents/policies/q_policy_test.py", line 269, in testActionWithinBoundsWithMasking
    self.assertTrue(np.all(action >= 0) and np.all(action < num_actions))
AssertionError: False is not true

----------------------------------------------------------------------
Ran 1 test in 0.115s

FAILED (failures=1)

When I don't specify an action mask by removing observation_and_action_constraint_splitter, it passes:

policy = q_policy.QPolicy(ts.time_step_spec(input_tensor_spec), action_spec, q_net)

(.venv) aziz@merlin:~/code/agents$ python -m unittest tf_agents.policies.q_policy_test.QPolicyTest.testActionWithinBoundsWithMasking
culprits: []
.
----------------------------------------------------------------------
Ran 1 test in 0.051s

OK

Is this behavior expected?

[tfboyd]

Ok, I feel silly. This is not a failing unit test. You created a unit test to illustrate the question regarding some behavior. Eugene will have to handle that question.