Goal-directed Learning to Fly

Proceedings of the Twentieth International Conference on Machine Learning

Volume

Proceedings of the Twentieth International Conference on Machine Learning

Contents

Learning to fly an aircraft is a complex task that requires the development of control skills and goal achievement strategies. This paper presents a behavioural cloning system that learns to successfully fly manoeuvres, in turbulence, of a realistic aircraft simulation. A hierarchical decomposition of the problem is employed where goal settings and the control skill to achieve them are learnt. The benefit of this goal-directed approach is demonstrated in the reuse of the learnt manoeuvres to a flight path that includes novel manoeuvres. The system is based on an error minimisation technique that benefits from the use of model tree learners. The model trees provide a compact and comprehensible representation of the underlying control skill and goal structure. The performance of the system was improved by compensating for human reaction times and goal anticipation. We conclude that our system addresses current limitations of behavioural cloning by producing robust, reusable and readable clones.