Investigation of Drive-Reinforcement Learning and Application of Learning toFlight Control

摘要

This report describes results obtained during a multiphase research programhaving the broad aim of investigating the application of learning systems to automatic control in general, and to flight control in particular. The first phase analyzed the drive-reinforcement learning paradigm and examined its application to automatic control, with mixed results. The second phase compared a number of alternative strategies for learning augmented control, and resulted in the conception of a new hybrid adaptive/learning control scheme. Subsequently, in the third phase, this hybrid control approach was more fully developed and applied to several nonlinear dynamical systems, including a cart-pole system, aeroelastic oscillator, and a three-degree-of-freedom aircraft. The fourth phase revisited drive-reinforcement learning from the point of view of optimal control and successfully applied a version embedded in the associative control process architecture to regulate an aeroelastic oscillator. The fifth phase examined the problem of learning augmented estimation, and resulted in the development of a preliminary estimation scheme consistent with the hybrid control approach. In the sixth and final phase, the hybrid control methodology was applied to a nonlinear, six-degree-of-freedom flight control problem, and then demonstrated via a challenging multiaxis maneuver. ACP Network, Drive-reinforcement, Reinforcement learning, Adaptive control, Hybrid control, Nonlinear control, Aircraft flight control, Learning control, Optimal control.