Building symbolic representations of intuitive real-time skills from performance data

Abstract

Real-time control skills are ordinarily tacit - their possessors cannot explicitly communicate them. But given sucient sampling of a trained expert's input-output behaviour, machine learning programs have been found capable of constructing rules which, when run as programs, deliver behaviours similar to those of the original exemplars. These 'clones' are in efect symbolic representations of subcognitive behaviours. After validation on simple pole-balancing tasks, the principles have been successfully generalized in flight-simulator experiments, both by Sammut and others at UNSW, and by Camacho at the Turing Institute. A flight plan switches control through a sequence of logically concurrent sets of reactive behaviours. Each set can be thought of as a committee of subpilots who are respectively specialized for rudder, elevators, rollers, thrust, etc. The chairman (the flight plan) knows only the mission sequence, and how to recognize the onset of each stage. This treatment is essentially that of the 'blackboard model', augmented by machine learning to extract subpilot behaviours (seventy-two behaviours in Camacho's auto-pilot for a simulated F-16 combat plane). A 'clean-up' effect, first noted in the pole-balancing phase of this enquiry, results in auto-pilots which fly the F-16 under tighter control than the human from whom the behavioural records were sampled.