Supervised Learning for Real-time Model Predictive Control Leveraging Multi-parametric Mixed Integer Quadratic Programming

This paper presents a novel supervised learning framework for real-time optimization of multi-parametric mixed-integer quadratic programming (mp-MIQP) problems. The framework utilizes a multi-layer perceptron (MLP) model to efficiently predict both continuous and binary control inputs while classifying the feasibility of the optimization problem. To address the computational burden of branch-and-bound methods and the memory limitations of explicit model predictive control (MPC), this framework learns optimal control inputs across diverse system states and prediction horizons through offline training. Comparative evaluations reveal that the proposed method significantly outperforms traditional approaches in computational speed and memory efficiency while maintaining high accuracy in approximating optimal solutions. These advancements position the framework as a transformative solution for real-time control applications, offering a robust and efficient alternative for online optimization in mp-MIQP problems.