Minimum Attention Control (MAC) in a Receding Horizon Framework with Applications

TL;DR

This paper introduces Minimum Attention Model Predictive Control (MAMPC), a receding horizon control method that minimizes input changes for efficient control, demonstrated through case studies on systems with different dynamics.

Contribution

It develops a novel MPC framework incorporating zero norm constraints for minimal input changes and provides an efficient alternating minimization algorithm for solving it.

Findings

Sparse control actions achieved with a tradeoff in tracking error.

Effective on systems with slow and fast dynamics.

Outperforms standard MPC in minimizing input changes.

Abstract

Minimum Attention Control (MAC) is a control technique that provides minimal input changes to meet the control objective. Mathematically, the zero norm of the input changes is used as a constraint for the given control objective and minimized with respect to the process dynamics. In this paper, along with the zero norm constraint, stage costs are also considered for reference tracking in a receding horizon framework. For this purpose, the optimal inputs of the previous horizons are also considered in the optimization problem of the current horizon. An alternating minimization algorithm is applied to solve the optimization problem (Minimum Attention Model Predictive Control (MAMPC)). The outer step of the optimization is a quadratic program, while the inner step, which solves for sparsity, has an analytical solution. The proposed algorithm is implemented on two case studies: a four-tank system with slow dynamics and a fuel cell stack with fast dynamics. A detailed comparative study of the proposed algorithm with standard MPC indicates sparse control actions with a tradeoff in the tracking error.