Changing the step size during evaluation about rl_reach HOT 4 CLOSED

ADDENDUM: I can manually increase --n-eval-steps to >100 steps, but then the evaluation resets after 100 timesteps.

from rl_reach.

The short answer is no. You should train and evaluate using the same environment. If you notice some oscillations at evaluation time, one solution is to increase the action space as you said by changing action_min and action_max in init.py. However you should then re-train.

Regarding your 2nd query, this is normal as the number of evaluation steps is independent of the max_episode_steps=100 in init.py. However it makes sense to set n-eval-steps as a multiple of max_episode_steps to evaluate with a finite number of episode. For example, if n-eval-steps = 2000 and max_episode_steps=100, you will use 20 full episodes to evaluate your trained policy. However it makes sense to evaluate over many episodes mainly if you use non-deterministic evaluation.

from rl_reach.

Thanks for the thorough explanation. As a short follow-up: Where do the action_min and action_max from the init.py overwrite the hard-coded action_min and action max in the environment?

from rl_reach.

action_min and action_max are used to define the boundaries of the action space here

from rl_reach.