Many current model reference adaptive control methods employ parametric adaptive elements in which the number of parameters are fixed a priori and the hyperparameters, such as the bandwidth, are predefined, often through expert judgment. As an alternative to these methods, a nonparametric model using Gaussian processes was recently proposed. Using Gaussian processes, it is possible to maintain constant coverage over the operating domain by adaptively selecting new kernel locations as well as adapt hyperparameters in an online setting to improve model prediction. In this work, the first extensive experimental flight results are presented using Gaussian process/model reference adaptive control. Experimental results show that Gaussian process/model reference adaptive control outperforms traditional model reference adaptive control methods that use radial basis function neural networks in terms of tracking error as well as transient behavior on trajectory following using a quadrotor. Results show an improvement of a factor of two to three over preexisting state-of-the-art methods. Additionally, many model reference adaptive control frameworks treat the adaptive element as being known exactly, and they do not incorporate certainty of the prior model into the control policy. In this paper, the notion of a Bayesian scaling factor is introduced that scales the adaptive element in order to incorporate the uncertainty of the prior model and current model confidence. The stability and convergence of using the Bayesian scaling factor in a closed loop is proven.
ASJC Scopus subject areas
- Aerospace Engineering
- Computer Science Applications
- Electrical and Electronic Engineering