Smoothening the model outputs

Hi all,

My PPO trained car is driving in lane but it keeps shaking left and right to stay at the center of the lane. How can i smoothen the control of the car? If PID, how can i integrate it to work with the RL model?

It’s kinda expected to have a jittery behaviour with RL, It also depends on how you formulate the actions and what is the rate of decision requests. You can also try imitation learning with demo files.

Hey,
have you tried smoothing your values in Update? ML-agent runs on Fixed Frames/ in fixedupdate, so you could try to interpolate there (e.g. the position, or output).

am i right to say that decreasing the rate of decision requests will lead to smoother actions?
for imitation learning, im assuming to use GAIL and set use_actions = true?