Allow the throttle ranges in joystick throttle

[Ezward]

Throttle is a difficult thing to train in the deep learning template presumably because it is very hard to have a consistent throttle when a human is driving. So we train on very inconsistent data and so the model is not good at predicting throttle. This is especially noticeable when the amount of data is not large.

We have a categorical model that buckets throttles and steering. But it always does that for both steering and throttle. Our linear steering model is very good, it's throttle that is the problem. So perhaps the solution is to have a model that is linear for steering but categorical for throttle.

Alternatively we could bucket the throttle provided by the game controller when the human is driving to collect the data. If we only had fast, medium, slow and stopped values from the game controller then we might be able to reduce the variability in the throttle and so end up with a linear model that does a better job with throttle. Since this would apply at the time the data is collected, the human would be driving with this in mind (as opposed to the categorical model which buckets after data collection; at training time) and so would tailor their driving style to it.