Neural Network Approaches for Model Predictive Control

Abstract: Model Predictive Control (MPC) is an optimization-based paradigm forfeedback control. The MPC relies on a dynamical model to make predictionsfor the future values of the controlled variables of the system. It then solvesa constrained optimization problem to calculate the optimal control actionthat minimizes the difference between the predicted values and the desiredor set values. One of the main limitations of the traditional MPC lies in thehigh computational cost resulting from solving the associated optimizationproblem online. Various offline strategies have been proposed to overcomethis, ranging from the explicit MPC (eMPC) to the recent learning-basedneural network approaches. This thesis investigates a framework for thetraining and evaluation of a neural network for learning to implement theMPC. As a part of the framework, a new approach for efficient generationof training data is proposed. Four different neural network structures arestudied; one of them is a black box network while the other three employMPC specific information. The networks are evaluated in terms of twodifferent performance metrics through experiments conducted on realistictwo-dimensional and four-dimensional systems. The experiments revealthat while using MPC specific structure in the neural networks resultsin performance gains when the training data is limited, all the networkstructures perform similarly as extensive training data is used. They furthershow that a recurrent neural network structure trained on both the state andcontrol trajectories of a family of MPCs is able to generalize to previouslyunseen MPC problems. The proposed methods in this thesis act as a firststep towards developing a coherent framework for characterization of learningapproaches in terms of both model validation and efficient training datageneration.