Multi input-multi output tank system data-driven model reference control

This paper suggests a model-free control approach for tuning nonlinear state feedback controllers to ensure model reference output tracking in an optimal control framework. An iterative Batch fitted Q (BFQ)-learning strategy uses two neural networks (NNs) to estimate the value function (critic) and the controller (actor). An initially stabilizing linear Virtual Reference Feedback Tuning (VRFT) controller learned from few input-output process samples is then used to collect significantly more input-state-output samples in a controlled constrained environment, by compensating for undesired process dynamics. This collected data is subsequently used to learn significantly superior nonlinear state feedback NN controllers for model reference output tracking using the proposed iterative BFQ-learning strategy. The mixed VRFT-BFQ learning approach is experimentally validated on the water level control of a multi input-multi output (MIMO) nonlinear constrained coupled two-tank system. Although the VRFT control is designed independently for each control channel and does not ensure decoupling, straightforward (MIMO) BFQ-learning proves good decoupling and ensures indirect linearization of the feedback MIMO control system.