michigan-traffic-lab / Dense-Deep-Reinforcement-Learning

This repo contains the code for paper "Dense reinforcement learning for safety validation of autonomous vehicles"
Other
304 stars 47 forks source link

how can i see various drl model activity #8

Open mikegithub123 opened 3 weeks ago

mikegithub123 commented 3 weeks ago

Hello, I was trying to get a vision of DRL's decision making, so I set mode to "D2RL",and run main.py, here are some issues: 1 Openning sumo-gui setting caused crash, I change sumo version to 1.20, and then I can watch vehicles in gui. 2 The activation of DRL model is very rare(can hardly see any vehicle turns blue), so I modified traffic generaion code, set all vehicles to be CF, still very rare. Then I reduced the distance between vehicles, and began to see DRL controlled vehicles.
3 So far as I can see, if controlled at relative back position of cav, the DRL vehicle tend to change to the same lane as cav, and tend to get a rear-end collision,but if controlled at relative front position of cav, the DRL vehicle performs no obvious action. So here are questions: 1 From the code I can see the DRL model controls at most one vehicle simultaneously, can this be modified so that more vehicles are controlled so the result will be more obvious? 2 what can I do to get vision of more various actions of DRL model? Thank you very much!