Using GRU models in the LPV-LFT framework

The use of neural networks is tending to become more widespread in control problems, but still suffers from a lack of connection with the classic tools of Automatic Control. In this paper, a Linear Parameter-Varying Linear Fractionnal Transformation (LPV-LFT) of a modified Gated Recurrent Unit (GRU) is proposed. A methodology is introduced to fully leverage this new formulation, thereby facilitating in perspective the construction of controllers with dedicated tools. Performance tests are conducted on both constrained and unconstrained GRU to analyze performance losses due to the assumptions of this LPV-LFT formulation.