Megainduction: A Test Flight

Abstract

This case study examines the application of Quinlan's C4.5 to the task of diagnosing a subsystem of NASA's Space Shuttle. Hundreds of thousands of training instances were available from simulator runs and real flight data. The trees produced are highly accurate, moderately small, and after being converted to production rules, were judged by the expert to be not only comprehensible and acceptable, but to contain new knowledge that might otherwise have remained undiscovered. The training set's huge size contributes to the high accuracy. The lack of noise turned out not to be so critical to accuracy, but learning time looks infeasible if extrapolated to a million examples. The complexity of the concept could also grow too large. We point to methods of removing these two stumbling blocks of current machine learning technology.