search

DongChen06 / MARL_CAVs

MARL for Autonomous Vehicles
216 stars 43 forks source link

Simulation and Emulation #44

Open zcysun opened 3 weeks ago

zcysun commented 3 weeks ago

Hello!

Currently, the simulation and emulation process involves fixed durations and numbers of vehicles, with each episode restarting from zero after it ends. Has the author considered fixing only the traffic scenario (such as a ramp junction) and generating vehicles at regular intervals, not starting and ending with an episode, but using simulation time for long-term training?

This approach seems to allow for a more comprehensive evaluation of the simulation process from multiple aspects, such as throughput, collision rate, and so on.

Best wishes!

DongChen06 commented 3 weeks ago

Hi, thanks for your suggestion. Yes, it is possible to do that, since our algorithm is based on offline buffer. In our scenarios, in each episode, the number of vehicles changes from 2-6 (easy), 4-8 (medium), and 7-11 (hard), which will not affect our algorithms. So to add a continuous traffic flow, you can modify the vehicle generation function in the merge_env.

zcysun commented 2 weeks ago

Hello! The training is evaluated at regular intervals. So, does the evaluation process give feedback on the training process? Or does the evaluation only read and collect relevant information and does not affect the training process?