This has been discussed a bit here:

http://code.google.com/p/ardupilot-mega/issues/detail?can=2&start=0&num=100&q=&colspec=ID%20Type%20Status%20Priority%20Milestone%20Owner%20Summary%20Attachments&groupby=&sort=&id=119

Self training PID is definitly the way of the future. 
I suppose in the mean time, this could be done in a crude manner to extend the flight envelope. Something like this:
-Tune the basic PID settings for flight speeds in or around approach speed in APM (the most critical phase of any flight) then set something in the code that says that if airspeed increases, reduce the rate by which the PID acts on the controls, some sort of percentage. Essentially, this would be like automatic dual rates for the autopilot.
For every airframe, we know roughly what the stall speed is and what speed it becomes a little too freaky to fly. With that in mind, it should be reasonably easy to induce a PID percentage or rate system in the programming to have a basic self adjusting system. Until some code Guru comes up with the real mccoy of self training PIDs, this could be a solution.
Just a thought.