Default observation values for non present features

CommonRoad-RL
1

Hello,

during evaluation of trained RL models I noticed that the observation values of features that are not present in a scenario (e.g. lane_based_rel_1) take constant values which are different from zero. These values are independent from the evaluated sceanrio, but dependent of the RL model used for the evaluation.
Edit: The default values for absent features represent 0.0 in the model’s normalized input space.

Can this be interpreted to characterize the trained model? E.g. as a bias, that the model has for each individual feature?

Best,
Annabel

2

Hi Annabel,

The lane-based surrounding observations contain information about six fixed neighbor vehicles, see Fig.2(a) in our paper. When there exists no neighbor vehicle at a certain position, we assume that there is a fake vehicle which the relative position to the ego vehicle is the sensing range and relative velocity is zero to consider occlusion. Using zero for relative position would mean that there is a collision with the agent happening.

Best,
Xiao