MAML Driver env not able to load SUMO configurations

RLlib
1

I am using a SUMO (Eclipse SUMO - Simulation of Urban MObility) based environment for my agent training. So far, the environment worked fine on PPO agent. Now, I modified the environment to be TaskSettableEnv to train MAML on it. The problem I am experiencing is that my environment and the SUMO configurations are getting loaded for the workers but not for the driver. MAML does not supporting turning off the environment creation for driver and the driver is not able to locate my configuration files for SUMO.

Why is the environment created on the driver process not able to load/locate the SUMO configuration files? Does it look for files in some different directory? How can I resolve this issue? In the logs below, I referred to the pid in order to infer that the driver environment is not getting created.

(pid=15768) 2021-11-17 03:11:47,126	INFO trainer.py:671 -- Tip: set framework=tfe or the --eager flag to enable TensorFlow eager execution
(pid=14564) Loading configuration ... done.
(pid=14564) 2021-11-17 03:11:52,628	INFO ppo_tf_policy.py:328 -- `vf_share_layers=True` in your model. Therefore, remember to tune the value of `vf_loss_coeff`!
(pid=14564) 2021-11-17 03:11:52,733	INFO tf_policy.py:154 -- TFPolicy (worker=1) running on CPU.
(pid=14564) 2021-11-17 03:11:52,758	INFO dynamic_tf_policy.py:472 -- Testing `compute_actions` w/ dummy batch.
(pid=14564) 2021-11-17 03:11:52,758	INFO tf_run_builder.py:87 -- Executing TF run without tracing. To dump TF timeline traces to disk, set the TF_TIMELINE_DIR environment variable.
(pid=14564) 2021-11-17 03:11:52,791	INFO dynamic_tf_policy.py:481 -- Adding extra-action-fetch `action_prob` to view-reqs.
(pid=14564) 2021-11-17 03:11:52,791	INFO dynamic_tf_policy.py:481 -- Adding extra-action-fetch `action_logp` to view-reqs.
(pid=14564) 2021-11-17 03:11:52,792	INFO dynamic_tf_policy.py:481 -- Adding extra-action-fetch `action_dist_inputs` to view-reqs.
(pid=14564) 2021-11-17 03:11:52,792	INFO dynamic_tf_policy.py:481 -- Adding extra-action-fetch `vf_preds` to view-reqs.
(pid=14564) 2021-11-17 03:11:52,792	INFO dynamic_tf_policy.py:488 -- Testing `postprocess_trajectory` w/ dummy batch.
(pid=14564) Loading configuration ... done.
(pid=15768) Error: Could not access configuration './envs/sumo_files_meta/sumo.sumocfg'.
(pid=15768) Quitting (on error).
(pid=15768) The current task id is: None
(pid=15768) C:\Program Files (x86)\Eclipse\Sumo\tools\traci\main.py:162: UserWarning: Could not connect to TraCI server using port 59149. Retrying with different port.
(pid=15768)   warnings.warn("Could not connect to TraCI server using port %s. Retrying with different port." % sumoPort)
(pid=15768) Error: Could not access configuration './envs/sumo_files_meta/sumo.sumocfg'.
(pid=15768) Quitting (on error).
(pid=15768) C:\Program Files (x86)\Eclipse\Sumo\tools\traci\main.py:162: UserWarning: Could not connect to TraCI server using port 59164. Retrying with different port.
(pid=15768)   warnings.warn("Could not connect to TraCI server using port %s. Retrying with different port." % sumoPort)
(pid=15768) Error: Could not access configuration './envs/sumo_files_meta/sumo.sumocfg'.
(pid=15768) Quitting (on error).

@sven1977

2

Update: I am able to make the training work now by setting the hard coded path of configuration file instead of relative path. So it is definitely the issue of Driver not been able to access the files from the relative path. Seems like a bug to me.

Now I face another problem. My training only works if I set the Hidden Layer size to be equal to my Observation Space size. This was not the case for PPO. Is the implementation of fcnet_hiddens configuration different for PPO and MAML?