[RLlib] Issue 8714: QMIX init error w/ tuple obs space. (#8936)

rllib/agents/qmix/qmix_policy.py

@@ -194,6 +194,7 @@ class QMixTorchPolicy(Policy):
             agent_obs_space = agent_obs_space.spaces["obs"]
         else:
             self.obs_size = _get_size(agent_obs_space)
+            self.env_global_state_shape = (self.obs_size, self.n_agents)
 
         self.model = ModelCatalog.get_model_v2(
             agent_obs_space,

rllib/tests/test_avail_actions_qmix.py

@@ -78,6 +78,8 @@ class TestAvailActionsQMix(unittest.TestCase):
         for _ in range(5):
             agent.train()  # OK if it doesn't trip the action assertion error
         assert agent.train()["episode_reward_mean"] == 21.0
+        agent.stop()
+        ray.shutdown()
 
 
 if __name__ == "__main__":