Reducing Memory Footprints in Explicit Model Predictive Control using Universal Numbers

Abstract

Abstract Explicit Model Predictive Control (MPC) is an effective alternative to reduce the on-line computational demand of traditional MPC. The idea of explicit MPC is to pre-compute the optimal MPC feedback law off-line and store it in a form of look-up table which is to be used in on-line phase. One of the main bottlenecks in an implementation of explicit MPC is memory required to store optimal solutions. This limit its applicability to systems with few states, small number of constraints, and short prediction horizons. In this paper, we present a novel way of reducing the memory footprint of explicit MPC solutions. The procedure is based on encoding all data (i.e., the critical regions and the feedback laws) as universal numbers (unums), which can be viewed as a memory-efficient extension of IEEE floating point standard. By doing so, we illustrate that the total memory footprint can be reduced by 80% without losing control accuracy. An additional advantage of proposed approach is, it can be applied on top of existing complexity reduction techniques.