Traffic light control algorithm and agent injection/deletion algorithm

[sfm-user-1]

Thanks for the excellent work developing and sharing MetaDrive!

Could you all describe more about the traffic light control algorithm and agent injection/deletion algorithms you use, if any? For example, how and where are agents spawned, and when are they deleted, and why? Do you use heuristics, and could you point me to the code where they are implemented? @pengzhenghao @QuanyiLi could you share that with me?

[pengzhenghao]

traffic light control algorithm We don't have any traffic light control on PG environments. In ScenarioNet environments, we replay the traffic light signal from data.

Agent injection/deletion for PG env is by procedual generation algorithm. For each map block there are some regions allowing spawn new agents and there are some predifined destinations in each block. So we can randomly select a spawn position and a destination for each agent. For scenario env, we replay data.