Neural Network Approaches for Model Predictive Control

Abstract: Model Predictive Control (MPC) is an optimization-based paradigm for feedback control. The MPC relies on a dynamical model to make predictions for the future values of the controlled variables of the system. It then solves a constrained optimization problem to calculate the optimal control action that minimizes the difference between the predicted values and the desired or set values. One of the main limitations of the traditional MPC lies in the high computational cost resulting from solving the associated optimization problem online. Various offline strategies have been proposed to overcome this, ranging from the explicit MPC (eMPC) to the recent learning-based neural network approaches. This thesis investigates a framework for the training and evaluation of a neural network for learning to implement the MPC. As a part of the framework, a new approach for efficient generation of training data is proposed. Four different neural network structures are studied; one of them is a black box network while the other three employ MPC specific information. The networks are evaluated in terms of two different performance metrics through experiments conducted on realistic two-dimensional and four-dimensional systems. The experiments reveal that while using MPC specific structure in the neural networks results in performance gains when the training data is limited, all the four network structures perform similarly as extensive training data is used. They further show that a recurrent neural network structure trained on both the state and control trajectories of a family of MPCs is able to generalize to previously unseen MPC problems. The proposed methods in this thesis act as a first step towards developing a coherent framework for characterization of learning approaches in terms of both model validation and efficient training data generation.