Added action calculation on not normalized observation space

[MarcelBruckner]

I added a second observer that has _NormalizationEnabled = False set on which we then calculate the action. Please have a look at the implementations if this is as you intended it.

There is now the problem that the actions are way out of the range of the action space defined in the load function. So your normalize function during load fails now. We have the possibilities to:

• Drop values out of [low, high]
• Normalize with the [min, max] values of the actions that are loaded

What do you think?

[ferenctorok]

Hey Marcel!

I've checked the things. I think everything looks all right except for the steering angle values. All the other values are within range (observations and accelerations too). I think the normalization of the actions should be all right, it is the same function that I also use for all the other normalizations. During tests there are 1202 values out of the 1600 that are out of range for the steering angles, and many of them are way out of range, not just a bit. So I hope I am right when I say, that there still is a bug in the steering angle calculation.

In the param files the steering angle is usually constrained to be within [-0.2, 0.2]. So an easy check would be to look at the values. My guess is that it will be easy to see, that most of them are out of range.

For later: The most reasonable solution in my opinion would be to cut out the points which still have out-of-range values. (Maybe there will still be some.) The agent is implemented so, that it can not produce bigger values than its constraints, so out of the range values would obviously be expert values. Not that it would make a big difference, but I guess still this would be the precise practice to follow.

What do you think?