search

Guojyjy / CoTV

Cooperative control for traffic light signals and connected autonomous vehicles using deep reinforcement learning
MIT License
62 stars 11 forks source link

visualize the trained model #6

Open Gavin-Tao opened 10 months ago

Gavin-Tao commented 10 months ago

Hi, thanks very much for your open source. When I run python ../flow/visualize/visualizer_rllib.py traind/model/path 200 --horizon 2000 --gen_emission, I found that the SUMO-GUI was opened a lot. In addition, I found that the SUMO-GUI only performs fixed-time signal control instead of using trained model. May I ask if you have encountered this problem?

What's more? May I ask if you use the trained output files or the visualizer_rllib output files to measure agents' performance? THe trained output files refer to the CoTV/output files when training the model. The visualizer_rllib output files refer to the Visualize/output files when run visualizer_rllib.py.

I am looking forward to hearing from you.

Guojyjy commented 6 months ago

About the fixed-time signal control, please refer to Issue #4, let me know if any further questions

Guojyjy commented 6 months ago

The experiment results in CoTV paper are averaged from the output files in 18 episodes after training convergence, covering the randomness of urban traffic and the stochasticity of DRL model operation.

After training has converged, either the output files when training the model or running visualizer_rllib.py can be used in the method evaluation, because visualizer_rllib.py actually restores the same trained model by checkpoint files:

https://github.com/Guojyjy/CoTV/blob/5ce1a95cc52d7a108294488a701025b488787f07/flow/flow/visualize/visualizer_rllib.py#L155