I’d have a really big problem with using learning/adaptive software (did someone say AI?) in safety-critical aerospace applications because such software deliberately fudges the boundaries of what it can and cannot do. That’s where learning occurs - at the edges of the envelope.
But, the edges of the envelope in aero software gets people killed.
I’m very skeptical, but not for the “crazy” reason. I distrust the “unexpected” solutions, too!
I’m in the certification side of aero software, BTW.
I was booked to give a talk at AUVSI’s Unmanned Systems Europe 2009 on that very subject; engineering with learning systems for applications that require rigid quality standards. We had critical events at the company then, so I couldn’t make it. I may yet write the paper though. It’s not so mysterious and difficult - but I’m hiding the reason ... for now. (Drum rolls, marketing hype, raising expectations, music plays, curtain slowly opens ...)
... This paper describes the new software architecture and discusses the potential of application. The current implementation can be installed on a wide range of autonomous systems, automatically locates sensors and actuators and builds its own system specific control programs. Local environment simulations are constructed from sensor input and used as a “robot's imagination” to adapt and solve problems. System behaviour can be extended by training in new environments and providing new challenges, by installing new “fitness functions” to drive learning, and by integrating application specific components developed “by hand.” Adaptation and field creation of new behaviour can be limited to accommodate various requirement levels for testing before use, from exploratory research and experimental development to the most rigid field-ready product quality control standards. We also expect the system to reduce development time and cost.