Small fix to rllib visualizer

[cathywu]

Checkpoint files (from RLlib experiments) used to be written to ~/ray_results/experiment_tag/Scenario/checkpoint_N/checkpoint_N.something, but now they seem to be written to ~/ray_results/experiment_tag/Scenario/checkpoint_N.something. This PR change reflects that change in directory structure, for when we try to visualize from the checkpoint.

@eugenevinitsky @fywu85 @AboudyKreidieh Could you confirm this?

[cathywu]

@eugenevinitsky @AboudyKreidieh I see that the Travis build is failing. Could there be a ray version difference in the Travis build?

Here is what a sample results directory looks like to me (no checkpoint_N subdirectories) with a clean install done today. But, the unit tests corresponding to the checkpoint locations are failing on Travis.

[AboudyKreidieh]

@cathywu I'm not sure about running these locally (haven't run rllib experiments since the new changes), but if you check in the data folder in tests they are under the subdirectory checkpoint_N, hence the failing tests.

[cathywu]

To clarify, I wanted to confirm whether the Travis build is using the latest version of flow-project/ray? I just did a clean install of everything yesterday, so it's unlikely that I'm using anything that's out-of-date. Is there a way we can add a check somewhere that notifies us when the Travis build is out of date with the SUMO / ray versions we are currently using?

[cathywu]

Update: I was mistaken above, the Travis builds are dynamically pulling from flow-project/ray, but from the ray_merge branch and not master. This discrepancy is the cause of the issue.

This PR was issued in accordance to incompatibility of the rllib_visualizer with flow-project/ray:master. Once flow-project/ray#2 is addressed, we can close this PR (without merge).

[eugenevinitsky]

Yeah, a fix for this is coming. We're behind Ray and should try to upgrade to 0.6.1

[AboudyKreidieh]

we've upgraded to ray 0.6.1 since, so I am assuming this is now resolved. If I am wrong, please feel free to reopen.