Giving a minimum green duration. about deeprl_signal_control HOT 3 CLOSED

This is tricky. 1) the minimum green duration must be a multiple of step size, i.e., control_interval_sec. 2) the easiest way to ensure this is to inset post-RL logic in env:step()

from deeprl_signal_control.

I understood the explanation for 1).

But I'm not sure how to do 2).

Could you please explain in more detail?

from deeprl_signal_control.

basically you do not restrict what RL learns, instead enforce your constraint when you simulate the step, by adjusting RL decision if it violates the constraint. e.g., RL:GGR->Actual:GGG if you need a minimum green duration of 3 MDP steps. Then RL will learn your enforcement automatically as part of env since the transition depends on actual actions.

from deeprl_signal_control.