Adaptive Reference-Augmented Predictive Control

We describe a novel adaptive non-linear model predictive controller which is based on the idea of neural-augmentation of reference elements, both at the level of the reduced model and at the level of the control action. The new methodology is primarily motivated by the desire to consistently incorporate existing legacy modeling and control techniques into an adaptive non-linear, yet real-time-capable, control framework. The proposed procedures are demonstrated in a virtual environment with the help of the classical model problem of the double inverted-pendulum, and with the more challenging reflexive control of an autonomous helicopter.