search

flow-project / flow

Computational framework for reinforcement learning in traffic control
MIT License
1.06k stars 372 forks source link

multiagent scenarios with different reward function unable to run #675

Open pnp91 opened 5 years ago

pnp91 commented 5 years ago

Hello, I have been working with Multi agents RL simulation but with different reward functions. It worked and unfortunately I had experienced a hard disk crash and had to reinstall everything from start. I have installed flow 0.3.0 version with ray 0.6.1 and Sumo

v0_31_0-812-g1d4338ab80

. I had previously worked with these same version set up. Unfortunately, multi agent files doesn't train and always throws a same error as shown in the screenshot. It trains for 10-15 training simulations and starts producing the error after collisions are started.

Bildschirmfoto von 2019-08-07 15-34-19

I am aware of the "SUMO Depart Time Issue" and had also replaced the line as mentioned in . But it does not help yet. I had checked for compatibility, version and reinstalled everything using anaconda, sumo as per Flow github installation setup:

Still the problem persists. where as single agent works well. I have been struggling with this issue since a week and still unable to find any solution or what exactly the problem is. Could some one of you help me how can I proceed and run the multi-agent files. I tried running multi agents of

Stabilizing the ring, cooperative merge and figure eight scenarios.

Versions : Ubuntu(Linux) 18.04 Flow : 0.3.0 Ray(rllib) : 0.6.1 Sumo : v0_31_0-812-g1d4338ab80

I would be glad to receive any help. Its been one week that I am stuck at the same problem. I hope someone can help me. Thanks a lot

eugenevinitsky commented 5 years ago

Thanks for flagging this!

Hmm, so we can look into the error and figure this out, but in terms of the fastest response to immediately make your errors go away, you could upgrade to Flow 0.4.0 which is compatible with a new version of SUMO that does not have this departure time bug anymore.

In the meantime we will try to reproduce and debug this.

pnp91 commented 5 years ago

Hi Vinitsky,

Thank you for your feedback. As per your suggestion I have tried with new version of flow and the problem still persists. Only improvement I could see was I am now able to run only Stabilizing the ring scenario but none of the other scenarios like figure eight, cooperative merge works. Is there any other suggestions or could you help me with understanding of what is missing, so that I can resimulate all my active files that were running before.

Thanks & Regards pnp91