[UncleOxidant]

Oh sure, I completely understand that. The point I was trying to make was that all of this flight control & simulation software is very specific to the characteristics of their particular kite. I'm not aware of anyone else in the industry making a kite that large and heavy - most of these companies are operating on a shoestring budget. Sure, it's possible that maybe some general ideas could be obtained from their autonomous control software - but they would need to be heavily adapted to a different kite with different flight characteristics.

We were starting to think about using reinforcement learning to develop an autonomous controller for our kite. That would have involved recording lots of real-world in-flight data for our particular kite which we were starting to do. Similar flight data for a different kite wouldn't have done us much good when it came to training as each kite has very different flight characteristics.

[tobinfricke]

We don't mention this at all in the technical report, but there was a lot of interest in applying reinforcement learning to the control problem, and there was some work done in this direction. (I was one of several who groaned whenever we were admonished to "just use machine learning." )

Ultimately we couldn't afford to dedicate much effort to the rather speculative RL approach (our team was small and we were always scrambling to meet our rather ambitious milestones - you know, the usual story :-) ). Instead we chose to continue to pursue tried-and-true controls techniques.