2. What I'd like to see is an exponential function for whether a unit's training decreases with a floor so that it wouldn't drop below a certain point. Maybe 20-25 for the floor? or at least make the % chance that training drops at that point low. That way units with high training would see decreases the fastest but units with moderate to low training wouldn't see large drops.
I think part of the current problems is that the function is exponential. Exponentials work in a way that the effect is really slow at first, and then at some threshold it gets very fast. That's bad as once you passed the threshold you're pretty much screwed, but before the threshold you don't get much effect.
A power law would be much better. Power laws are scale invariant. Roughly, this means that every day will be worst than the next by a similar amount, with no set threshold.
