Adaptive Control Using Gaussian-Process with Model Reference Generative Network

Abstract

In this paper, we present a new architecture for Gaussian Processes Model Reference Adaptive Control (GP-MRAC) trained using a generative network. GP-MRAC is a successful method for achieving global performance in the systems enabling adaptive control. GP-MRAC can handle a broader set of uncertainties without requiring apriori knowledge of the domain of operation. However, existing GP-MRAC work requires estimates of the state-derivate, and this is a primary limitation in the implementation of the controller. In this paper, we alleviate this major limitation by creating Model reference adaptive framework as Generative Network (MRGeN). Our contribution is a generative network architecture for learning Gaussian model to predict system uncertainties without having to estimate the state derivatives while ensuring that the system stability properties are unaffected. We retain the nonparametric nature of the controller by sharing the kernels between GP's and MRGeN, ensuring global performance and stability guarantees. GP-MRGeN can also be viewed as a method of baseline policy transfers, with applications in Reinforcement Learning.